Focusing on practical applications of machine learning (ML) within the Amazon Web Services ecosystem. The content systematically covers the exam syllabus, starting with ML fundamentals like modeling pipelines, supervised and unsupervised learning, and data splitting strategies to prevent overfitting and underfitting. It then details various AWS services for AI/ML, including Amazon Rekognition for image/video analysis, Amazon Polly for text-to-speech, Amazon Transcribe for speech-to-text, and Amazon Comprehend for natural language processing (NLP), alongside storage solutions like Amazon S3, RDS, and Redshift. The guide also explains data preparation and transformation techniques, such as handling missing values, outliers, and unbalanced datasets, and explores different ML algorithms (e.g., linear regression, XGBoost, K-means) as well as their evaluation and optimization through metrics like precision, recall, and hyperparameter tuning using Amazon SageMaker.
You can listen and download our episodes for free on more than 10 different platforms:
https://linktr.ee/cyber_security_summary
Get the Book now from Amazon:
https://www.amazon.com/Certified-Machine-Learning-Specialty-Certification/dp/1800569009?&linkCode=ll1&tag=cvthunderx-20&linkId=ec5d3390e2fb431864ff90c4f68df62c&language=en_US&ref_=as_li_ss_tl
Discover our free courses in tech and cybersecurity, Start learning today:
https://linktr.ee/cybercode_academy
You can listen and download our episodes for free on more than 10 different platforms:
https://linktr.ee/cyber_security_summary
Get the Book now from Amazon:
https://www.amazon.com/Certified-Machine-Learning-Specialty-Certification/dp/1800569009?&linkCode=ll1&tag=cvthunderx-20&linkId=ec5d3390e2fb431864ff90c4f68df62c&language=en_US&ref_=as_li_ss_tl
Discover our free courses in tech and cybersecurity, Start learning today:
https://linktr.ee/cybercode_academy
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:00):
Welcome back to the deep dive, where we unpack complex
topics and bring you the essential insights. Today we're navigating
the exciting world of machine learning on Amazon Web Services.
Our mission for this deep dive is really to distill
the core concepts of machine learning, walk you through its
end to end life cycle, the whole process, and also
highlight how AWS services provide a well powerful toolkit for
(00:24):
every single stage. We're drawing our insights from the AWS
Certified Machine Learning Specialty MLSC zero one certification guide, trying
to pull out those aha moments and strategic takeaways. The
goal is to make you feel truly well informed, whether
you're a strategizing for a meeting or just really curious
about this field.
Speaker 2 (00:42):
Yeah, and what's truly valuable in this guide and what
we'll focus on today is its ability to break down
these complex mL ideas into actionable knowledge. We're going to
try and give you a clear roadmap basically from foundational
definitions right through to practical AWOS applications, showing you how
you might build truly intelligent solutions.
Speaker 1 (00:59):
Okay, let's dig in to this.
Speaker 2 (01:00):
Then.
Speaker 1 (01:00):
The guide starts by clarifying the relationship between artificial intelligence
machine learning and deep learning. I like the analogy they use.
Think of it like a set of nested Russian dolls exactly.
Speaker 2 (01:11):
So at the outermost layer you've got artificial intelligence AI.
That's the really broad field, aiming to create machines that
can do tasks mimicking human intelligence. Then moving inward, machine
learning mL is a key subset of AI. This is
where systems learn from data. They identify patterns and make
predictions without being explicitly programmed. It's about learning from experience,
(01:35):
observing adapting.
Speaker 1 (01:36):
Okay, learning from data, not rules precisely.
Speaker 2 (01:39):
And then at the very core you have deep learning DL.
That's an even more specialized subset of mL. Deep learning
uses these multi layered structures you've probably heard of them,
deep neural networks. They solve highly complex problems. They're powering
a lot of the state of the art stuff we
see today, like language translation or facial recognition.
Speaker 1 (01:57):
So what this hierarchy really means for us, for you listening,
is that we're witnessing this incredible evolution. It's fueled by
well more computing power and just vast amounts of data
being available now, and AI applications are becoming more powerful,
more accessible and really applicable across almost every industry.
Speaker 2 (02:15):
And when these systems learn, they generally fall into three
main approaches, three ways of learning. The first is supervised learning.
This relies on labeled data. So imagine you have a
data set where every example has an answer already attached.
That's your labeled.
Speaker 1 (02:30):
Data, right, like inputs and the correct outputs exactly.
Speaker 2 (02:33):
So. One common use for supervised learning is classification. Here
the model predicts a category or class. For instance, the
guy talks about classifying financial transactions. Is this fraudulent or legitimate?
Based on features like amount, time of day, that sort.
Speaker 1 (02:48):
Of thing, okay, putting things into buckets?
Speaker 2 (02:50):
Yeah? And the other key type is regression. The goal
here is to predict a continuous numerical value. This could
be forecasting sales figures for the next quarter maybe, or
predicting the obstable price for.
Speaker 1 (03:01):
A product, got it, predicting a number.
Speaker 2 (03:03):
Then there's unsupervised learning. This works with unlabeled data, so
no answer is provided beforehand. Here, this system tries to
find hidden patterns or structures on its own.
Speaker 1 (03:14):
Ah, okay, so finding patterns we didn't know.
Speaker 2 (03:16):
We're there exactly a great example of this is clustering
you group similar data points together. Think about segmenting your
customer base based on their purchasing behavior. You know, to
understand different.
Speaker 1 (03:27):
Market segments, right, finding natural groupings.
Speaker 2 (03:30):
And finally, we have reinforcement learning. This is where a
system learns by interacting with an environment. It gets rewards
for good decisions and well penalties for poor ones. It's
a bit like how we learn through trial and error.
The guide mentions an example like an automated call center
agent learning the best path to resolve customer queries by
(03:51):
getting rewarded for good recommendations.
Speaker 1 (03:53):
Interesting learning by doing essentially So, this next point seems
crucial because it's about how we actually use these The
approach you choose, supervised, unsupervised, or reinforcement, it totally depends
on your data and the problem you're trying to.
Speaker 2 (04:07):
Solve, right, absolutely, it's fundamental. Do you have clearly labeled examples?
Supervised is likely your path? Are you looking for hidden
groups in just raw data? Unsupervised? Is it about learning
through interaction and feedback? Reinforcement The data and the goal
dictate the method.
Speaker 1 (04:25):
Makes sense. Now, building effective mL models isn't just about
picking one of those algorithms. It's a structured process. The
guide highlights something called crisp DM, the cross industry's standard
process for data mining, as a blueprint for this.
Speaker 2 (04:38):
Yeah, chris DM is really widely used. It provides a clear,
iterative framework with six key phases. It starts with business understanding.
This is all about clearly defining your project objectives, your
success criteria, potential risks. It sounds obvious, but honestly, this
is where many projects can go wrong if the problem
isn't nailed down.
Speaker 1 (04:55):
Precisely right, knowing what you're actually trying to achieve.
Speaker 2 (04:58):
Then data understanding. This involves collecting, describing, exploring, checking the
quality of your raw data. Data scientists need to be
well super skeptical here, look for every nuance. Then comes
data preparation and this is often the most time consuming phase. Really.
It involves selecting, cleaning, transforming, formatting the data for your chosen.
Speaker 1 (05:20):
Algorithm, Okay, getting the data ready.
Speaker 2 (05:22):
Following that is modeling. Here you select the appropriate algorithm,
design your tests, approach and train the model. You need
to distinguish between parameters, which are learned from the data itself,
and hyper parameters, which are like knobs you turn to
control the learning process. The fifth phase is evaluation. You
review the model's performance against those initial business success criteria you.
Speaker 1 (05:42):
Defined and if it's not good enough.
Speaker 2 (05:45):
That's the key. mL is iterative. It's a scientific process.
If your model isn't cutting it, you loop back, maybe
you tune those hyper parameters, maybe you need more data,
maybe you even need to rethink the business problem itself. Finally, deployment,
getting your model into reduction. This involves creating pipelines for
continuous training and inference and setting up monitoring to catch
(06:06):
model drift. All drift, Yeah, that's what a model's performance
degrades over time because the real world data or patterns change,
So you need to monitor and potentially retrain. And you know,
if we connect this back to the AWS certification, the
four domains covered in the exam data engineering, exploratory, data analysis, modeling,
and mL OPS, right, they really map quite directly to
(06:27):
these CRISP DM stages. It's complete life cycle.
Speaker 1 (06:30):
Okay, that framework makes a lot of sense. Now, you
mentioned data preparation is often the most time consuming part.
The guide really stresses this too. It's the absolute foundation
for any good model. Get the data wrong, and well
nothing else.
Speaker 2 (06:41):
Matters much absolutely garbage in, garbage out. As they say,
A critical first step is understanding your feature types, the
kind of beta you have. So you've got numerical data.
This could be discrete like countable items, number of clicks maybe,
or continuous measurements with potentially infinite values like temperature or.
Speaker 1 (06:59):
Price, numbers, screen or continuous.
Speaker 2 (07:01):
Got it. Then you have categorical data. This describes qualities
or labels. It can be nominal labels without any inherent order,
like colors or types of products, or ordinal labels that
do have a meaningful order like low, medium, high, or
education levels.
Speaker 1 (07:16):
Okay, categories with or without an order.
Speaker 2 (07:18):
Right, and categorical data, especially nominal, usually can't be fed
directly into most algorithms. It needs transforming into numbers. For example,
for that nominal data without order, like countries, we often
use one hot encoding. This creates a new binary column
a zero or one for each category. It avoids accidentally
implying that, say, country three is somehow greater than country two.
Speaker 1 (07:39):
Ah avoids creating a false order exactly, whereas for ordinal
data like those education levels, ordinal encoding preserves that inherent sequence.
Speaker 2 (07:48):
Now the crucial rule here in this trips people up
sometimes is that any encoder you create must be fitted
only on your training data. Then you use that same
fitted encoder to transform your teches data and any new
production data. You never refit on test data that introduces bias.
Speaker 1 (08:05):
Okay, fit on train, transform on tests.
Speaker 2 (08:07):
Got it now. For numerical features, you often need to
adjust their scale. Data normalization, for instance, might scale data
to arrange between zero and one. This is really vital
for algorithms that are sensitive to the magnitude of numbers,
like neural networks or caneurous.
Speaker 1 (08:21):
Neighbors, so they don't overweight big numbers precisely.
Speaker 2 (08:24):
Alternatively, data standardization transforms data to have a mean of
zero and a standard deviation of one. This is fantastic
for identifying outliers, for example, and for features that are
skewed think income distributions often bunched up at one end.
Logarithmic and power transformations like the box Cox method can
make them more symmetrical, more like a Bell curve, and
(08:45):
that often significantly improves the performance of many algorithms like
linear regression.
Speaker 1 (08:50):
Wow, lots of ways to wrangle the data. What about
problems like missing values?
Speaker 2 (08:55):
Yeah, that's a common one. First, you have to try
and understand why they're missing. Is it ran them or
is there a pattern. Options range from just listwise deletion
discarding rows or columns with missing data, but be careful
you might lose valuable information, or imputation where you replace
missing values. Simple imputation might use the mean or the median,
which is less sensitive to outliers, or the mode for
(09:16):
categorical data, but you can get more sophisticated even using
other mL models to predict what the missing values should be.
Speaker 1 (09:23):
Okay, and outliers those weird data points.
Speaker 2 (09:26):
So another common hurdle. Outliers are data points significantly different
from the rest. They can dramatically skew your model's understanding,
like pulling a regression line way off course. Tools like
z scores or visualizing with box plots help detect them.
Once found, you might remove them or maybe just flag
them so your model knows they're unusual.
Speaker 1 (09:46):
Makes sense, And what if the data is like really unbalanced.
You mentioned fraud detection earlier.
Speaker 2 (09:51):
Right, Unbalanced data sets very common. Say only one percent
of your transactions are actually fraudulent. Your model might just
learn to always predict not fraud, because that's accurate ninety
nine percent of the time, but it misses the important cases.
So to address this, you can tune your algorithm, maybe
tell to pay more attention to the rare class using
something like a class weight hyperparameter. Or you can resample
(10:12):
your data. Either undersample the majority class just use fewer
examples of not fraud, or oversample the minority class. A
popular technique for oversampling is SMO and a synthetic minority
over sampling technique. It intelligently creates new synthetic examples of
the rare class to help balance things.
Speaker 1 (10:29):
Out smot okay, creating fake but plausible examples kind.
Speaker 2 (10:34):
Of yeah, based on the characteristics of the existing minority examples,
and finally preparing text data for mL or natural language
processing NLP. This has evolved a lot. Older methods like
bag of Words BOW just count how often words appear simple,
but loses context. More advanced techniques like word embedding, used
(10:56):
in models like word two, VEK or glove represent words
as dense numerical vectors. What's fascinating here is these vectors
capture semantic meaning. Words with similar meanings end up closer
together in this multi dimensional space, so.
Speaker 1 (11:09):
The model understands relationships between words in.
Speaker 2 (11:11):
A mathematical sense. Yes, it captures context and meaning much
better than just counting.
Speaker 1 (11:16):
That's a really thorough look at data prep. It's clear
its critical and well often complex. But all this meticulously
prepared information needs a robust place to live. You need
to store it somewhere, and on AWS. That journey often
begins with S three, Right, our digital warehouse, where do
we store all this data? For mL?
Speaker 2 (11:33):
You're absolutely right. The storage choice is fundamental. Amazon S
three Simple Storage Service is very often the starting point
and the core its object storage, known for its incredible durability,
designed for eleven nine's durability, which is just astronomical protection
against data loss. It's highly scalable. You store objects your
(11:53):
files basically within these things called buckets, which are specific
to an AWS region, and S three offers different storage classes.
This lets you optimize costs based on how frequently you
need to access the data. Data you access rarely can
go into cheaper, colder storage. Plus, it has robust access
control and encryption options to keep everything secure. OK.
Speaker 1 (12:12):
S three for scalable, durable object storage, what about more
structured data like traditional databases?
Speaker 2 (12:18):
For that, Amazon Relational Database Service RDS is the managed service.
It supports popular engines like Mycycle, Postgress, Goal, Oracle, etc.
A key feature for reliability is multi easy deployments. This
automatically creates a synchronous standby copy of your database in
a different availability zone, so if one AZ has an issue,
it fails over automatically.
Speaker 1 (12:36):
Great for high availability, so it keeps running even if
there's an outage in one place exactly.
Speaker 2 (12:41):
And for scaling read performance, especially for applications that do
a lot of reading, you can use read replicas. These
are asynchronously replicated copies of your main database. You can
point your read heavy traffic to them. You can even
place them in different regions for global reach. This directly
impacts your RPO recovery point objective how much data you
might lose an RTO recovery time objective how fast you recover.
(13:04):
Multi asy and read replicas help you achieve low RPO
and RTO.
Speaker 1 (13:08):
Makes sense availability and read scaling.
Speaker 2 (13:10):
HM and beyond S three and rds AWS has specialized
stores too. Amazon Redshift is a data warehouse optimized for
analyzing massive data sets using SQL and Amazon DynamoDB is
a fully managed no SQL database the key value in
document data where you need super fast, flexible access at
really any scale.
Speaker 1 (13:30):
Okay, so a whole range of options. The key takeaway
here seems to be it's not just about storing data,
it's about choosing the right storage for the right kind
of data, getting that optimal balance of availability, performance, security,
and cost for your specific mL use.
Speaker 2 (13:44):
Case, precisely matching the tool to the job.
Speaker 1 (13:47):
So once our data is carefully stored and prepped, we
often need to process it further, maybe transform it in
bulk or analyze streams of it. The guide walks us
through a WUS services for both batch processing and real
time stuff.
Speaker 2 (13:59):
Yeah, large scale data transformation and movement like etlxtract transform
load AWS. Glue is a really powerful, fully managed service.
It's a secret Sauce is the data catalog. You can
automatically crawl your data sources, figure out the schema, detect changes,
and make it all queriable. Then glues ETL jobs, which
usually run on a patchy spark, do the heavy lifting
(14:22):
of the actual data transformation, maybe copying and cleaning data
from S three into redshift for example.
Speaker 1 (14:27):
So Glue handles the whole ETL pipeline.
Speaker 2 (14:29):
Pretty much in a serverlest way. Now, if you just
want a query data that's already sitting in S three
without moving or transforming it first, Amazon Athena is amazing
for this. It's serverless, interactive use standard SQL to query
data directly in S three across various formats CSV, json, parquet, ORC,
no infrastructure to manage. Is incredibly fast for ad hoc
analysis or quick.
Speaker 1 (14:50):
Exploration schema onread right, you define the structure as.
Speaker 2 (14:53):
You query it exactly. Now, for processing real time streaming data,
we turn to Amazon Kinesis Visais data streams can capture
and store huge amounts of data per second from loads
of sources website clicks, IoT sensors, financial transactions. You can
then build applications to process this stream in real time.
Then there's Kinesis Data fire Hose. This is a fully
(15:14):
managed service that takes that streaming data and automatically loads
it into destinations like S three, redshift or analytics services.
It can even transform the data on the fly using
AWS Lambda before delivering it.
Speaker 1 (15:25):
So fire Hose is more about getting the stream into
storage or other services easily.
Speaker 2 (15:30):
Yeah, simplifies the delivery part. And what about getting data
from your own data centers into AWS. AWS Storage Gateway
connects your on premises software appliances to cloud storage using
standard file or block protocols. For really massive data transfers
where the Internet is too slow, you have the AWS
snow family. These are physical devices like Snowball Edge which
(15:51):
is like a ruggedized suitcase computer, or even Snowmobile, a
whole shipping container. You load data onto them locally, ship
them to AWS and they upload it securely, much faster
for petabytes a truck.
Speaker 1 (16:01):
Full of data literally pretty much.
Speaker 2 (16:04):
And AWS Data Sinc. Is great for ongoing online data
transfer between your on premises storage and AWS services like
S three or EFS. Finally, for those really big computation
heavy batch jobs, things that might take hours or days
or need massive resources beyond what Lander offers, Aws Bachil
lets you schedule and run these efficiently. It manages the
(16:24):
job queues, provisions the right compute resources like EC two instances,
and scales automatically.
Speaker 1 (16:30):
Okay. This really covers the spectrum, from analyzing static data
with Athena and glue to handling real time streams with kinesis,
and even moving massive data sets physically. AWS seems to
have a tool for almost every data processing need.
Speaker 2 (16:43):
It's a very comprehensive set of services.
Speaker 1 (16:45):
Now, before we dive headfirst into coding raw algorithms, the
guide makes a point of highlighting aws's out of the
box AI services. These seem designed to make advanced mL
accessible even if you're not a deep learning expert. Right,
no model building recques exactly.
Speaker 2 (17:01):
These are pre trained managed services. You use them via
simple API calls. They bring sophisticated AI capabilities directly into
your applications with minimal fuss. For example, Amazon Recognition provides
powerful visual analysis. It can detect objects, people, faces, texts
and images, and videos, even sentiment analysis on faces. Amazon
Polly converts text into remarkably lifelike speech, loads of voices languages,
(17:24):
great for accessibility or creating voice interfaces.
Speaker 1 (17:27):
Polly Speaks and Recognition Ce's right.
Speaker 2 (17:29):
And Amazon transcribed as the opposite of poly. It converts
speech into text, excellent for transcribing audio, video calls, generating captions.
It supports custom vocabularies too, for better accuracy and specific domains.
Amazon comprehend digs into unstructured text I think customer reviews, emails,
social media feeds. It pulls out insights like sentiment positive, negative, neutral,
(17:50):
key phrases, entities, even topics.
Speaker 1 (17:52):
So comprehend understands text.
Speaker 2 (17:55):
Amazon Translate provides high quality, real time language translation between languages.
Amazon TExtract is really interesting. It goes beyond basic ocr
optical character recognition. It understands the structure of documents, so
it can extract data not just as raw text, but
specifically from forms and tables, preserving their layout and relationships.
Super useful for document.
Speaker 1 (18:16):
Processing while understanding forms and tables not just text.
Speaker 2 (18:20):
Yeah, and finally, Amazon Lex this is the engine that
powers Amazon Alexa. It lets you build sophisticated conversational interfaces chatbots,
voice spots using natural language understanding NLU and automatic speech recognition. ASR.
You define the user's goals, intense the information needed, slots
and sample phrases, utterances, and LEX handles the complex conversation flow.
Speaker 1 (18:45):
Okay, that's an incredible menu of ready to use AI
really lowers the barrier to entry, But it begs the
question for you listening, how do you decide when should
you use these powerful pre built tools versus actually diving
in and building a custom mL model from scratch.
Speaker 2 (19:01):
That's a really important strategic decision and the answer often
comes down to specificity and control. For common, well defined
tasks like general translation, sentiment analysis, standard object recognition, and images,
these managed services are often the fastest, easiest, and most
cost effective path. They're pre trained by AWS on massive
data sets, so you benefit from that expertise with minimal
(19:22):
development effort. You don't need deep mL knowledge to integrate
them via APIs.
Speaker 1 (19:27):
So use them for the standard stuff.
Speaker 2 (19:28):
Generally, yes, However, if your problem is highly specialized, maybe
involves unique data types not covered by the services, or
if you need fine grained control over the model architecture
or the training process or the specific performance trade offs.
That's when building a custom model, probably using a platform
like Amazon sage Maker becomes the better choice. It gives
you full flexibility, but requires more mL expertise and effort.
Speaker 1 (19:52):
Got it. Use managed services for speed and common tasks,
build custom for unique needs and control. Okay, now let's
go deeper into the custom model, building into the heart
of mL, the algorithms themselves. The guide outlines aws's built
in algorithms available in sage Maker, which are often optimized
for the AWS environment. But first, maybe a quick word
on ensemble models. The guide mentions these are pretty powerful.
Speaker 2 (20:13):
Yeah. Ensemble methods are a really important concept. The idea
is to combine multiple individual mL models to get better
predictive performance than any single model could achieve on its own.
Two main types are bagging, think bootstrap aggregating. Like in
a random forest algorithm, You train many models, usually decision trees,
independently on different random samples of your data, and then
(20:36):
you average their predictions for regression or take a majority
vote for classification. It helps reduce variants.
Speaker 1 (20:43):
So wisdom of the crowd applied to models.
Speaker 2 (20:45):
Kinda yeah. The other main type is boosting. Here models
are trained sequentially. Each new model focuses on correcting the
errors made by the previous ones. It builds a strong
predictor Iteratively, algorithms like ATTA boost or the very popular
XG boost uses approach. Boosting often leads to very high accuracy,
but you need to be careful about overfitting.
Speaker 1 (21:05):
Okay, bagging is parallel boosting a sequential makes sense? So
what are some of the key built in algorithms sage
Maker offers, for say, supervised learning.
Speaker 2 (21:13):
Right for supervised tasks with labeled data, sage Maker has
several optimized algorithms. The linear learner algorithm is a good
starting point. It's versatile handling with regression, predicting numbers and
classification predicted categories. It's great for understanding linear relationships and
includes options like L one and L two regularization to
prevent overfitting and even perform some automatic feature selection. Then
(21:38):
there's XG boost. As we mentioned, it's a gradi at
boosting algorithm, incredibly popular and often wins data science competitions,
especially with structured tabular data. Sage Maker has a highly
optimized version.
Speaker 1 (21:49):
XG boost seems like a go to for many problems.
Speaker 2 (21:52):
It often is for unsupervised learning finding patterns in unlabeled data.
K means is a classic clustering algorithm. You tell how
many clusters you want to find, and it groups your
data points based on similarity typically distance. Great for customer
segmentation or finding archetypes. Random cut Forest RCF is specifically
designed for anomaly detection. It builds a collection of random
(22:13):
trees and identifies data points that are easily isolated. These
are likely anomalies, good for fraud or outlier.
Speaker 1 (22:18):
Detection, finding the odd ones out exactly.
Speaker 2 (22:21):
And principal component analysis PCA. This is a fundamental technique
for dimensionality reduction. If you have lots and lots of features,
PCA can transform them into a smaller set of uncorrelated
principal components that capture most of the original information. This
helps simplify models, reduce noise, sometimes improve performance, and even
makes high dimensional data easier to visualize.
Speaker 1 (22:42):
Reducing complexity while keeping the important info.
Speaker 2 (22:46):
That's the goal. Sage Maker also has specialized algorithms like
deeper for time series forecasting using sophisticated recurrent neural networks.
For text analysis, there's blazing text, which is optimized for
both text classification and generating work word embeddings like word
twvec very quickly on large data sets, And of course
a suite of algorithms for image processing image classification that's
(23:07):
the main object object detection find multiple objects in drawboxes,
and semantic segmentation classify every pixel in the image.
Speaker 1 (23:14):
Wow, so a really broad set of tools. What about
data formats? Do these algorithms just take CSV files?
Speaker 2 (23:20):
Many can take text CSSV Yes. For supervised learning, the
convention is usually the target variable in the first calumn
no head or row. However, for peak performance and efficiency,
especially with large data sets, many stage Maker built in
algorithms prefer an optimized binary format called recordio protobuff. This
format allows for something called pipe mode where data is
(23:42):
streamed directly from S three to the training instance without
needing to download it all first. It saves time and
disk space. Uh.
Speaker 1 (23:50):
Recordio protobuff for speed and streaming. This is a really
comprehensive toolkit. It's clear AWS provides these highly optimized tools
for almost any mL task. It lets you, the user
focus more on framing the problem and interpreting results, rather
than getting totally bogged down in the low level infrastructure
or algorithm implementation.
Speaker 2 (24:09):
That's definitely the aim of a managed service like sage Maker.
Speaker 1 (24:12):
Okay, so we've built these potentially incredible models using these algorithms,
but how do we actually know if they're any good?
How do we evaluate them? It's not just about hitting
run and hoping for the best.
Speaker 2 (24:21):
Right, absolutely not. Evaluation is critical. It's not just about
getting a single accuracy number. It's about understanding how your
model performs, its strengths, its weaknesses, and whether it actually
meets the business need. Evaluation metrics are crucial for documenting performance,
comparing different models or different versions of the same model,
tracking them over time and production, and importantly for detecting
(24:45):
that model drift we talked about earlier, when performance degrades
metrics tell you it's time to retrain or investigate.
Speaker 1 (24:51):
So it's about ongoing quality control too.
Speaker 2 (24:54):
Definitely for classification models, the ones predicting categories like fraud
not fraud, or spam not spam. The confusion matrix is fundamental.
It's a simple table that breaks down predictions versus actual outcomes.
You get four key numbers. True positives TP correctly predicted positive,
said fraud, was fraud. True negative PN correctly predicted negatives
not fraud, wasn't fraud. False positives FP incorrectly predicted positive,
(25:15):
said fraud, but wasn't. This is a Type I error
of false alarm false negatives. FN incorrectly predicted negative, said
not fraud, but was fraud. This is a type two
error of missed detection.
Speaker 1 (25:24):
TP tn fp FN. Okay, the four outcomes.
Speaker 2 (25:27):
Right, and from this matrix we derive the most common
classification metrics accuracy TP plus TN divided by the total
overall correctness. But careful. It can be really misleading if
your data set is unbalanced like that ninety nine percent
not fraud. Example, recall or sensitivity tp TP plus fn.
This measures how well the model finds all the positive cases.
(25:50):
High recall is crucial when missing a positive is bad
reg missing a disease diagnosis. Precision or positive predictive value
tp TP plus fp this measures how often the model
is correct when it does predict positive High precision is
key when false alarms are costly, for example, marking important
emails as spam.
Speaker 1 (26:07):
Recall finds them all. Precision avoids false alarms. A trade off.
Speaker 2 (26:11):
Often, yes, there's usually a trade off between precision and recall.
The f one score is the harmonic mean of the two,
providing a single score that balances both. Useful when both
precision and recall are important.
Speaker 1 (26:21):
Okay, and what about those curves you see like ROC.
Speaker 2 (26:23):
Right evaluation curves help visualize that trade off across different
decision thresholds. The Precision Recall PR curve plots precision versus recall.
It's particularly useful for imbalanced data sets as it focuses
directly on the performance on the minority class. The ROC
curve receiver operating characteristic plots the true positive rate, which
is just recall, against the false positive rate FP FP
(26:46):
plus TN. It's commonly used for more balanced data sets.
The area under the curve AUC summarizes the curve into
a single.
Speaker 1 (26:52):
Number PR curve for imbalance, ROC for balance. Good tip.
What about regression models? The ones predicting numbers.
Speaker 2 (26:58):
Different metrics there, since we're not dealing with classes. Common
ones include MAE mean absolute error, the average of the
absolute differences between predictions and actual values. Simple intuitive units
MS mean squared error, the average of the squared differences.
This penalizes larger errors much more heavily than smaller ones.
(27:19):
RMS root means squared error the square root of MS.
This brings the metric back into the same units as
your target variable, making it easier to interpret while still
penalizing large errors. RMSE is probably the most common regression metric,
and MAP mean absolute percentage error calculates the error as
an average percentage of the actual values. Very intuitive for
things like sales forecasting.
Speaker 1 (27:40):
Okay, MA rmsc AMAPE for regression. That's a lot of metrics.
If you know, if you're listening and looking at these,
what's maybe one piece of advice you'd give about picking
the right metric for your specific project.
Speaker 2 (27:52):
That's a great question. The single most important thing is
to deeply understand your business goal and the cost of
different types of errors. Don't just default to accuracy because
it sounds good. Ask yourself what's worse a false positive
or a false negative. In medical diagnosis, missing a disease,
a false negative could be catastrophic, so you'd optimize for recall.
(28:13):
In filtering spam, marking a crucial email as spam, a
false positive is highly annoying, so you'd prioritize precision. The
context dictates the metric, Always tie it back to the
real world impact.
Speaker 1 (28:25):
Connect the metric to the business impact excellent advice.
Speaker 2 (28:28):
So once we can measure our models effectively using these metrics,
the next logical step is optimization, specifically hyperparameter tuning. Remember
those knobs that control the learning process. Finding the best
combination of those settings for your specific data is crucial.
The goal isn't just to get a model that performs
well in the data it was trained on. It's to
get a model that generalizes well to new unseen data.
(28:51):
We want to minimize both bias oversimplification and variance overfitting.
Speaker 1 (28:55):
Finding that sweet spot between too simple and too complex.
Speaker 2 (28:58):
Exactly there's several techniques for this hyper parameter search. Grid
search is the most basic. You define a grid of
possible values for each hyper parameter, and it literally tests
every single combination very thorough but can be incredibly slow
and computationally expensive, especially if you have many hyper parameters
or wide ranges. A more efficient approach is random search.
(29:21):
Instead of trying every combination, it randomly samples combinations from
your defined search space. Surprisingly, it often finds very good
or even optimal hyper parameters, much faster than grid search.
Speaker 1 (29:32):
Randomly trying things can be faster.
Speaker 2 (29:34):
Often yes, because not all hyper parameters are equally important.
Random search explores the space more broadly quicker. For even
more intelligence, there's Bayesian optimization. This method learns from past evaluations.
It builds a probability model of how hyper parameters relate
to performance, and uses it to intelligently choose the next
set of hyper parameters to try. Focusing on promising regions
(29:55):
of the search space. It can converge on optimal settings
much much faster, especially for complex models.
Speaker 1 (30:01):
Subaesian learns as it goes.
Speaker 2 (30:02):
Precisely, it's a smarter search.
Speaker 1 (30:04):
The core idea here, then, is that tuning isn't just
a one shot deal. It's an empirical process. You combine
these smart search strategies with solid evaluation techniques, often using
things like cross validation to build robust models, Models that
don't just memorize the training data, but actually perform well
out there in the real world.
Speaker 2 (30:24):
That's the name of the game generalization.
Speaker 1 (30:26):
Okay, bringing this all together. Now we've talked about the
life cycle, data algorithms, evaluation optimization. The guide clearly points
to Amazon stage Maker as the central workbench, the main
hub for doing all this on AWS.
Speaker 2 (30:38):
Yes, sage Maker is designed to be that integrated environment
for the entire mL workflow. It's a fully managed service
aiming to simplify each step. It provides notebook instances. These
are basically managed Jupiter notebooks running on EC two instances.
They're great for data exploration, cleaning, preprocessing, and generally orchestrating
your mL pipeline. For the heavy listing of training, stage
(31:00):
Maker provides dedicated, optimized training instances. You choose the instance
type based on your needs, submit your training code, and
Stagemaker handles provisioning, execution, and tearing down the resources. And
once your model is trained, sage Maker offers endpoint instances
for deploying it and getting real time predictions via a
simple API call. It handles scaling and availability for you.
Speaker 1 (31:21):
Notebooks for exploring training, instances for building endpoints for predicting.
Speaker 2 (31:26):
That's a good summary, and sage Maker also has managed
services specifically for hyper parameter tuning jobs. You de find
your hyper parameters ranges, the metric you want to optimize,
and SageMaker automatically runs the search using strategies like basing, optimization,
random search, or grid search, keeping track of the best performing.
Speaker 1 (31:41):
Job automating that tuning process we just discussed.
Speaker 2 (31:44):
Exactly when you choose instance types for these sage Maker components,
they're all based on EC two instances, often with mL prefixes.
You need to consider your workload. There are general purpose
M family, Compute optimized CE family, Memory optimized OUR family,
and portantly GPU enabled instances like the P and G families.
GPUs are essential for accelerating deep learning training, which involves
(32:08):
massive matrix calculations. The choice really depends on your data size,
algorithm complexity, budget, and how fast you need.
Speaker 1 (32:15):
Results matching the hardware to the mL task. What about security?
Keeping notebooks and data private.
Speaker 2 (32:20):
Security is built in you can launch sage Maker components
like notebook instances or training jobs within your own private
VPC virtual private cloud. This gives you fine grain control
over network access. You can restrict internet access, connect securely
to your on premises data sources, use security groups and
network acls. SageMaker also supports network isolation for training and
(32:41):
inference containers, preventing them from making unauthorized outbound network calls.
Speaker 1 (32:45):
So you can lock it down pretty tightly.
Speaker 2 (32:47):
Absolutely and while sage Maker provides that integrated environment, you
can also orchestrate mL workflows using other AWS services, sometimes
in combination. AWS lamb to functions, ervalless Event driven compute
functions are great for automating parts of the pipeline. For example,
an S three upload event could trigger a lamb to
function to do some initial data preprocessing or validation. For
(33:10):
more complex multi step workflows, AWS step functions is fantastic.
It lets you define your workflow as a visual state machine.
You can sequence and coordinate calls to Lambda functions, glue jobs, SageMaker,
training jobs, manual approval steps, pretty much any AWS service.
It's great for managing long running distributed processes with built
in error handling.
Speaker 1 (33:30):
And retries step functions for orchestrating the whole flow.
Speaker 2 (33:33):
Yeah, and if we connect this back to that bigger picture,
sage Maker, often combined with services like Glue, Lambda and
step functions, really aims to provide a managed, scalable, and
secure environment for that entire Cristium life cycle. We started
with from understanding the business need and preparing data in
notebooks all the way through training, tuning, deploying, and monitoring
(33:54):
models in production. It's designed to be end to end.
Speaker 1 (33:57):
And there you have it. Wow, what an insightful deep
dive that was. We started with the Very Foundation's AIMLDL.
We explored that critical CRISPM life cycle. We understood the
frankly immense importance of data preparation and the variety of
AWS storage options. Then we delved into the specific AWSAI
application services, those ready made tools, and also the powerful
(34:20):
built in algorithms within sage Maker. And finally we saw
how crucial evaluation and optimization are and how sage Maker
and other services help operationalize it. All You've just gained,
I think, a really valuable shortcut to being well informed
about the whole landscape of machine learning on AWS. Extracting
hopefully incredible value from what's normally a pretty dense technical guide.
Speaker 2 (34:40):
And maybe this raises an important final question for you,
the listener, to consider. Given the power you've just heard
about in these integrated AWS services, from intelligent data processing
and diverse storage to that comprehensive suite of AI and
mL tools, how might you reimagine a current data analysis task,
or maybe an automation challenge in your own work? How
(35:01):
could you potentially transform it into an intelligent, scalable mL
solution using some of these capabilities.
Speaker 1 (35:08):
That's a great thought to leave everyone with. How can
you apply this? Thanks for diving deep with us today.
Until next time, keep exploring, keep learning, and stay curious.
Welcome back to the deep dive, where we unpack complex
topics and bring you the essential insights. Today we're navigating
the exciting world of machine learning on Amazon Web Services.
Our mission for this deep dive is really to distill
the core concepts of machine learning, walk you through its
end to end life cycle, the whole process, and also
highlight how AWS services provide a well powerful toolkit for
(00:24):
every single stage. We're drawing our insights from the AWS
Certified Machine Learning Specialty MLSC zero one certification guide, trying
to pull out those aha moments and strategic takeaways. The
goal is to make you feel truly well informed, whether
you're a strategizing for a meeting or just really curious
about this field.
Speaker 2 (00:42):
Yeah, and what's truly valuable in this guide and what
we'll focus on today is its ability to break down
these complex mL ideas into actionable knowledge. We're going to
try and give you a clear roadmap basically from foundational
definitions right through to practical AWOS applications, showing you how
you might build truly intelligent solutions.
Speaker 1 (00:59):
Okay, let's dig in to this.
Speaker 2 (01:00):
Then.
Speaker 1 (01:00):
The guide starts by clarifying the relationship between artificial intelligence
machine learning and deep learning. I like the analogy they use.
Think of it like a set of nested Russian dolls exactly.
Speaker 2 (01:11):
So at the outermost layer you've got artificial intelligence AI.
That's the really broad field, aiming to create machines that
can do tasks mimicking human intelligence. Then moving inward, machine
learning mL is a key subset of AI. This is
where systems learn from data. They identify patterns and make
predictions without being explicitly programmed. It's about learning from experience,
(01:35):
observing adapting.
Speaker 1 (01:36):
Okay, learning from data, not rules precisely.
Speaker 2 (01:39):
And then at the very core you have deep learning DL.
That's an even more specialized subset of mL. Deep learning
uses these multi layered structures you've probably heard of them,
deep neural networks. They solve highly complex problems. They're powering
a lot of the state of the art stuff we
see today, like language translation or facial recognition.
Speaker 1 (01:57):
So what this hierarchy really means for us, for you listening,
is that we're witnessing this incredible evolution. It's fueled by
well more computing power and just vast amounts of data
being available now, and AI applications are becoming more powerful,
more accessible and really applicable across almost every industry.
Speaker 2 (02:15):
And when these systems learn, they generally fall into three
main approaches, three ways of learning. The first is supervised learning.
This relies on labeled data. So imagine you have a
data set where every example has an answer already attached.
That's your labeled.
Speaker 1 (02:30):
Data, right, like inputs and the correct outputs exactly.
Speaker 2 (02:33):
So. One common use for supervised learning is classification. Here
the model predicts a category or class. For instance, the
guy talks about classifying financial transactions. Is this fraudulent or legitimate?
Based on features like amount, time of day, that sort.
Speaker 1 (02:48):
Of thing, okay, putting things into buckets?
Speaker 2 (02:50):
Yeah? And the other key type is regression. The goal
here is to predict a continuous numerical value. This could
be forecasting sales figures for the next quarter maybe, or
predicting the obstable price for.
Speaker 1 (03:01):
A product, got it, predicting a number.
Speaker 2 (03:03):
Then there's unsupervised learning. This works with unlabeled data, so
no answer is provided beforehand. Here, this system tries to
find hidden patterns or structures on its own.
Speaker 1 (03:14):
Ah, okay, so finding patterns we didn't know.
Speaker 2 (03:16):
We're there exactly a great example of this is clustering
you group similar data points together. Think about segmenting your
customer base based on their purchasing behavior. You know, to
understand different.
Speaker 1 (03:27):
Market segments, right, finding natural groupings.
Speaker 2 (03:30):
And finally, we have reinforcement learning. This is where a
system learns by interacting with an environment. It gets rewards
for good decisions and well penalties for poor ones. It's
a bit like how we learn through trial and error.
The guide mentions an example like an automated call center
agent learning the best path to resolve customer queries by
(03:51):
getting rewarded for good recommendations.
Speaker 1 (03:53):
Interesting learning by doing essentially So, this next point seems
crucial because it's about how we actually use these The
approach you choose, supervised, unsupervised, or reinforcement, it totally depends
on your data and the problem you're trying to.
Speaker 2 (04:07):
Solve, right, absolutely, it's fundamental. Do you have clearly labeled examples?
Supervised is likely your path? Are you looking for hidden
groups in just raw data? Unsupervised? Is it about learning
through interaction and feedback? Reinforcement The data and the goal
dictate the method.
Speaker 1 (04:25):
Makes sense. Now, building effective mL models isn't just about
picking one of those algorithms. It's a structured process. The
guide highlights something called crisp DM, the cross industry's standard
process for data mining, as a blueprint for this.
Speaker 2 (04:38):
Yeah, chris DM is really widely used. It provides a clear,
iterative framework with six key phases. It starts with business understanding.
This is all about clearly defining your project objectives, your
success criteria, potential risks. It sounds obvious, but honestly, this
is where many projects can go wrong if the problem
isn't nailed down.
Speaker 1 (04:55):
Precisely right, knowing what you're actually trying to achieve.
Speaker 2 (04:58):
Then data understanding. This involves collecting, describing, exploring, checking the
quality of your raw data. Data scientists need to be
well super skeptical here, look for every nuance. Then comes
data preparation and this is often the most time consuming phase. Really.
It involves selecting, cleaning, transforming, formatting the data for your chosen.
Speaker 1 (05:20):
Algorithm, Okay, getting the data ready.
Speaker 2 (05:22):
Following that is modeling. Here you select the appropriate algorithm,
design your tests, approach and train the model. You need
to distinguish between parameters, which are learned from the data itself,
and hyper parameters, which are like knobs you turn to
control the learning process. The fifth phase is evaluation. You
review the model's performance against those initial business success criteria you.
Speaker 1 (05:42):
Defined and if it's not good enough.
Speaker 2 (05:45):
That's the key. mL is iterative. It's a scientific process.
If your model isn't cutting it, you loop back, maybe
you tune those hyper parameters, maybe you need more data,
maybe you even need to rethink the business problem itself. Finally, deployment,
getting your model into reduction. This involves creating pipelines for
continuous training and inference and setting up monitoring to catch
(06:06):
model drift. All drift, Yeah, that's what a model's performance
degrades over time because the real world data or patterns change,
So you need to monitor and potentially retrain. And you know,
if we connect this back to the AWS certification, the
four domains covered in the exam data engineering, exploratory, data analysis, modeling,
and mL OPS, right, they really map quite directly to
(06:27):
these CRISP DM stages. It's complete life cycle.
Speaker 1 (06:30):
Okay, that framework makes a lot of sense. Now, you
mentioned data preparation is often the most time consuming part.
The guide really stresses this too. It's the absolute foundation
for any good model. Get the data wrong, and well
nothing else.
Speaker 2 (06:41):
Matters much absolutely garbage in, garbage out. As they say,
A critical first step is understanding your feature types, the
kind of beta you have. So you've got numerical data.
This could be discrete like countable items, number of clicks maybe,
or continuous measurements with potentially infinite values like temperature or.
Speaker 1 (06:59):
Price, numbers, screen or continuous.
Speaker 2 (07:01):
Got it. Then you have categorical data. This describes qualities
or labels. It can be nominal labels without any inherent order,
like colors or types of products, or ordinal labels that
do have a meaningful order like low, medium, high, or
education levels.
Speaker 1 (07:16):
Okay, categories with or without an order.
Speaker 2 (07:18):
Right, and categorical data, especially nominal, usually can't be fed
directly into most algorithms. It needs transforming into numbers. For example,
for that nominal data without order, like countries, we often
use one hot encoding. This creates a new binary column
a zero or one for each category. It avoids accidentally
implying that, say, country three is somehow greater than country two.
Speaker 1 (07:39):
Ah avoids creating a false order exactly, whereas for ordinal
data like those education levels, ordinal encoding preserves that inherent sequence.
Speaker 2 (07:48):
Now the crucial rule here in this trips people up
sometimes is that any encoder you create must be fitted
only on your training data. Then you use that same
fitted encoder to transform your teches data and any new
production data. You never refit on test data that introduces bias.
Speaker 1 (08:05):
Okay, fit on train, transform on tests.
Speaker 2 (08:07):
Got it now. For numerical features, you often need to
adjust their scale. Data normalization, for instance, might scale data
to arrange between zero and one. This is really vital
for algorithms that are sensitive to the magnitude of numbers,
like neural networks or caneurous.
Speaker 1 (08:21):
Neighbors, so they don't overweight big numbers precisely.
Speaker 2 (08:24):
Alternatively, data standardization transforms data to have a mean of
zero and a standard deviation of one. This is fantastic
for identifying outliers, for example, and for features that are
skewed think income distributions often bunched up at one end.
Logarithmic and power transformations like the box Cox method can
make them more symmetrical, more like a Bell curve, and
(08:45):
that often significantly improves the performance of many algorithms like
linear regression.
Speaker 1 (08:50):
Wow, lots of ways to wrangle the data. What about
problems like missing values?
Speaker 2 (08:55):
Yeah, that's a common one. First, you have to try
and understand why they're missing. Is it ran them or
is there a pattern. Options range from just listwise deletion
discarding rows or columns with missing data, but be careful
you might lose valuable information, or imputation where you replace
missing values. Simple imputation might use the mean or the median,
which is less sensitive to outliers, or the mode for
(09:16):
categorical data, but you can get more sophisticated even using
other mL models to predict what the missing values should be.
Speaker 1 (09:23):
Okay, and outliers those weird data points.
Speaker 2 (09:26):
So another common hurdle. Outliers are data points significantly different
from the rest. They can dramatically skew your model's understanding,
like pulling a regression line way off course. Tools like
z scores or visualizing with box plots help detect them.
Once found, you might remove them or maybe just flag
them so your model knows they're unusual.
Speaker 1 (09:46):
Makes sense, And what if the data is like really unbalanced.
You mentioned fraud detection earlier.
Speaker 2 (09:51):
Right, Unbalanced data sets very common. Say only one percent
of your transactions are actually fraudulent. Your model might just
learn to always predict not fraud, because that's accurate ninety
nine percent of the time, but it misses the important cases.
So to address this, you can tune your algorithm, maybe
tell to pay more attention to the rare class using
something like a class weight hyperparameter. Or you can resample
(10:12):
your data. Either undersample the majority class just use fewer
examples of not fraud, or oversample the minority class. A
popular technique for oversampling is SMO and a synthetic minority
over sampling technique. It intelligently creates new synthetic examples of
the rare class to help balance things.
Speaker 1 (10:29):
Out smot okay, creating fake but plausible examples kind.
Speaker 2 (10:34):
Of yeah, based on the characteristics of the existing minority examples,
and finally preparing text data for mL or natural language
processing NLP. This has evolved a lot. Older methods like
bag of Words BOW just count how often words appear simple,
but loses context. More advanced techniques like word embedding, used
(10:56):
in models like word two, VEK or glove represent words
as dense numerical vectors. What's fascinating here is these vectors
capture semantic meaning. Words with similar meanings end up closer
together in this multi dimensional space, so.
Speaker 1 (11:09):
The model understands relationships between words in.
Speaker 2 (11:11):
A mathematical sense. Yes, it captures context and meaning much
better than just counting.
Speaker 1 (11:16):
That's a really thorough look at data prep. It's clear
its critical and well often complex. But all this meticulously
prepared information needs a robust place to live. You need
to store it somewhere, and on AWS. That journey often
begins with S three, Right, our digital warehouse, where do
we store all this data? For mL?
Speaker 2 (11:33):
You're absolutely right. The storage choice is fundamental. Amazon S
three Simple Storage Service is very often the starting point
and the core its object storage, known for its incredible durability,
designed for eleven nine's durability, which is just astronomical protection
against data loss. It's highly scalable. You store objects your
(11:53):
files basically within these things called buckets, which are specific
to an AWS region, and S three offers different storage classes.
This lets you optimize costs based on how frequently you
need to access the data. Data you access rarely can
go into cheaper, colder storage. Plus, it has robust access
control and encryption options to keep everything secure. OK.
Speaker 1 (12:12):
S three for scalable, durable object storage, what about more
structured data like traditional databases?
Speaker 2 (12:18):
For that, Amazon Relational Database Service RDS is the managed service.
It supports popular engines like Mycycle, Postgress, Goal, Oracle, etc.
A key feature for reliability is multi easy deployments. This
automatically creates a synchronous standby copy of your database in
a different availability zone, so if one AZ has an issue,
it fails over automatically.
Speaker 1 (12:36):
Great for high availability, so it keeps running even if
there's an outage in one place exactly.
Speaker 2 (12:41):
And for scaling read performance, especially for applications that do
a lot of reading, you can use read replicas. These
are asynchronously replicated copies of your main database. You can
point your read heavy traffic to them. You can even
place them in different regions for global reach. This directly
impacts your RPO recovery point objective how much data you
might lose an RTO recovery time objective how fast you recover.
(13:04):
Multi asy and read replicas help you achieve low RPO
and RTO.
Speaker 1 (13:08):
Makes sense availability and read scaling.
Speaker 2 (13:10):
HM and beyond S three and rds AWS has specialized
stores too. Amazon Redshift is a data warehouse optimized for
analyzing massive data sets using SQL and Amazon DynamoDB is
a fully managed no SQL database the key value in
document data where you need super fast, flexible access at
really any scale.
Speaker 1 (13:30):
Okay, so a whole range of options. The key takeaway
here seems to be it's not just about storing data,
it's about choosing the right storage for the right kind
of data, getting that optimal balance of availability, performance, security,
and cost for your specific mL use.
Speaker 2 (13:44):
Case, precisely matching the tool to the job.
Speaker 1 (13:47):
So once our data is carefully stored and prepped, we
often need to process it further, maybe transform it in
bulk or analyze streams of it. The guide walks us
through a WUS services for both batch processing and real
time stuff.
Speaker 2 (13:59):
Yeah, large scale data transformation and movement like etlxtract transform
load AWS. Glue is a really powerful, fully managed service.
It's a secret Sauce is the data catalog. You can
automatically crawl your data sources, figure out the schema, detect changes,
and make it all queriable. Then glues ETL jobs, which
usually run on a patchy spark, do the heavy lifting
(14:22):
of the actual data transformation, maybe copying and cleaning data
from S three into redshift for example.
Speaker 1 (14:27):
So Glue handles the whole ETL pipeline.
Speaker 2 (14:29):
Pretty much in a serverlest way. Now, if you just
want a query data that's already sitting in S three
without moving or transforming it first, Amazon Athena is amazing
for this. It's serverless, interactive use standard SQL to query
data directly in S three across various formats CSV, json, parquet, ORC,
no infrastructure to manage. Is incredibly fast for ad hoc
analysis or quick.
Speaker 1 (14:50):
Exploration schema onread right, you define the structure as.
Speaker 2 (14:53):
You query it exactly. Now, for processing real time streaming data,
we turn to Amazon Kinesis Visais data streams can capture
and store huge amounts of data per second from loads
of sources website clicks, IoT sensors, financial transactions. You can
then build applications to process this stream in real time.
Then there's Kinesis Data fire Hose. This is a fully
(15:14):
managed service that takes that streaming data and automatically loads
it into destinations like S three, redshift or analytics services.
It can even transform the data on the fly using
AWS Lambda before delivering it.
Speaker 1 (15:25):
So fire Hose is more about getting the stream into
storage or other services easily.
Speaker 2 (15:30):
Yeah, simplifies the delivery part. And what about getting data
from your own data centers into AWS. AWS Storage Gateway
connects your on premises software appliances to cloud storage using
standard file or block protocols. For really massive data transfers
where the Internet is too slow, you have the AWS
snow family. These are physical devices like Snowball Edge which
(15:51):
is like a ruggedized suitcase computer, or even Snowmobile, a
whole shipping container. You load data onto them locally, ship
them to AWS and they upload it securely, much faster
for petabytes a truck.
Speaker 1 (16:01):
Full of data literally pretty much.
Speaker 2 (16:04):
And AWS Data Sinc. Is great for ongoing online data
transfer between your on premises storage and AWS services like
S three or EFS. Finally, for those really big computation
heavy batch jobs, things that might take hours or days
or need massive resources beyond what Lander offers, Aws Bachil
lets you schedule and run these efficiently. It manages the
(16:24):
job queues, provisions the right compute resources like EC two instances,
and scales automatically.
Speaker 1 (16:30):
Okay. This really covers the spectrum, from analyzing static data
with Athena and glue to handling real time streams with kinesis,
and even moving massive data sets physically. AWS seems to
have a tool for almost every data processing need.
Speaker 2 (16:43):
It's a very comprehensive set of services.
Speaker 1 (16:45):
Now, before we dive headfirst into coding raw algorithms, the
guide makes a point of highlighting aws's out of the
box AI services. These seem designed to make advanced mL
accessible even if you're not a deep learning expert. Right,
no model building recques exactly.
Speaker 2 (17:01):
These are pre trained managed services. You use them via
simple API calls. They bring sophisticated AI capabilities directly into
your applications with minimal fuss. For example, Amazon Recognition provides
powerful visual analysis. It can detect objects, people, faces, texts
and images, and videos, even sentiment analysis on faces. Amazon
Polly converts text into remarkably lifelike speech, loads of voices languages,
(17:24):
great for accessibility or creating voice interfaces.
Speaker 1 (17:27):
Polly Speaks and Recognition Ce's right.
Speaker 2 (17:29):
And Amazon transcribed as the opposite of poly. It converts
speech into text, excellent for transcribing audio, video calls, generating captions.
It supports custom vocabularies too, for better accuracy and specific domains.
Amazon comprehend digs into unstructured text I think customer reviews, emails,
social media feeds. It pulls out insights like sentiment positive, negative, neutral,
(17:50):
key phrases, entities, even topics.
Speaker 1 (17:52):
So comprehend understands text.
Speaker 2 (17:55):
Amazon Translate provides high quality, real time language translation between languages.
Amazon TExtract is really interesting. It goes beyond basic ocr
optical character recognition. It understands the structure of documents, so
it can extract data not just as raw text, but
specifically from forms and tables, preserving their layout and relationships.
Super useful for document.
Speaker 1 (18:16):
Processing while understanding forms and tables not just text.
Speaker 2 (18:20):
Yeah, and finally, Amazon Lex this is the engine that
powers Amazon Alexa. It lets you build sophisticated conversational interfaces chatbots,
voice spots using natural language understanding NLU and automatic speech recognition. ASR.
You define the user's goals, intense the information needed, slots
and sample phrases, utterances, and LEX handles the complex conversation flow.
Speaker 1 (18:45):
Okay, that's an incredible menu of ready to use AI
really lowers the barrier to entry, But it begs the
question for you listening, how do you decide when should
you use these powerful pre built tools versus actually diving
in and building a custom mL model from scratch.
Speaker 2 (19:01):
That's a really important strategic decision and the answer often
comes down to specificity and control. For common, well defined
tasks like general translation, sentiment analysis, standard object recognition, and images,
these managed services are often the fastest, easiest, and most
cost effective path. They're pre trained by AWS on massive
data sets, so you benefit from that expertise with minimal
(19:22):
development effort. You don't need deep mL knowledge to integrate
them via APIs.
Speaker 1 (19:27):
So use them for the standard stuff.
Speaker 2 (19:28):
Generally, yes, However, if your problem is highly specialized, maybe
involves unique data types not covered by the services, or
if you need fine grained control over the model architecture
or the training process or the specific performance trade offs.
That's when building a custom model, probably using a platform
like Amazon sage Maker becomes the better choice. It gives
you full flexibility, but requires more mL expertise and effort.
Speaker 1 (19:52):
Got it. Use managed services for speed and common tasks,
build custom for unique needs and control. Okay, now let's
go deeper into the custom model, building into the heart
of mL, the algorithms themselves. The guide outlines aws's built
in algorithms available in sage Maker, which are often optimized
for the AWS environment. But first, maybe a quick word
on ensemble models. The guide mentions these are pretty powerful.
Speaker 2 (20:13):
Yeah. Ensemble methods are a really important concept. The idea
is to combine multiple individual mL models to get better
predictive performance than any single model could achieve on its own.
Two main types are bagging, think bootstrap aggregating. Like in
a random forest algorithm, You train many models, usually decision trees,
independently on different random samples of your data, and then
(20:36):
you average their predictions for regression or take a majority
vote for classification. It helps reduce variants.
Speaker 1 (20:43):
So wisdom of the crowd applied to models.
Speaker 2 (20:45):
Kinda yeah. The other main type is boosting. Here models
are trained sequentially. Each new model focuses on correcting the
errors made by the previous ones. It builds a strong
predictor Iteratively, algorithms like ATTA boost or the very popular
XG boost uses approach. Boosting often leads to very high accuracy,
but you need to be careful about overfitting.
Speaker 1 (21:05):
Okay, bagging is parallel boosting a sequential makes sense? So
what are some of the key built in algorithms sage
Maker offers, for say, supervised learning.
Speaker 2 (21:13):
Right for supervised tasks with labeled data, sage Maker has
several optimized algorithms. The linear learner algorithm is a good
starting point. It's versatile handling with regression, predicting numbers and
classification predicted categories. It's great for understanding linear relationships and
includes options like L one and L two regularization to
prevent overfitting and even perform some automatic feature selection. Then
(21:38):
there's XG boost. As we mentioned, it's a gradi at
boosting algorithm, incredibly popular and often wins data science competitions,
especially with structured tabular data. Sage Maker has a highly
optimized version.
Speaker 1 (21:49):
XG boost seems like a go to for many problems.
Speaker 2 (21:52):
It often is for unsupervised learning finding patterns in unlabeled data.
K means is a classic clustering algorithm. You tell how
many clusters you want to find, and it groups your
data points based on similarity typically distance. Great for customer
segmentation or finding archetypes. Random cut Forest RCF is specifically
designed for anomaly detection. It builds a collection of random
(22:13):
trees and identifies data points that are easily isolated. These
are likely anomalies, good for fraud or outlier.
Speaker 1 (22:18):
Detection, finding the odd ones out exactly.
Speaker 2 (22:21):
And principal component analysis PCA. This is a fundamental technique
for dimensionality reduction. If you have lots and lots of features,
PCA can transform them into a smaller set of uncorrelated
principal components that capture most of the original information. This
helps simplify models, reduce noise, sometimes improve performance, and even
makes high dimensional data easier to visualize.
Speaker 1 (22:42):
Reducing complexity while keeping the important info.
Speaker 2 (22:46):
That's the goal. Sage Maker also has specialized algorithms like
deeper for time series forecasting using sophisticated recurrent neural networks.
For text analysis, there's blazing text, which is optimized for
both text classification and generating work word embeddings like word
twvec very quickly on large data sets, And of course
a suite of algorithms for image processing image classification that's
(23:07):
the main object object detection find multiple objects in drawboxes,
and semantic segmentation classify every pixel in the image.
Speaker 1 (23:14):
Wow, so a really broad set of tools. What about
data formats? Do these algorithms just take CSV files?
Speaker 2 (23:20):
Many can take text CSSV Yes. For supervised learning, the
convention is usually the target variable in the first calumn
no head or row. However, for peak performance and efficiency,
especially with large data sets, many stage Maker built in
algorithms prefer an optimized binary format called recordio protobuff. This
format allows for something called pipe mode where data is
(23:42):
streamed directly from S three to the training instance without
needing to download it all first. It saves time and
disk space. Uh.
Speaker 1 (23:50):
Recordio protobuff for speed and streaming. This is a really
comprehensive toolkit. It's clear AWS provides these highly optimized tools
for almost any mL task. It lets you, the user
focus more on framing the problem and interpreting results, rather
than getting totally bogged down in the low level infrastructure
or algorithm implementation.
Speaker 2 (24:09):
That's definitely the aim of a managed service like sage Maker.
Speaker 1 (24:12):
Okay, so we've built these potentially incredible models using these algorithms,
but how do we actually know if they're any good?
How do we evaluate them? It's not just about hitting
run and hoping for the best.
Speaker 2 (24:21):
Right, absolutely not. Evaluation is critical. It's not just about
getting a single accuracy number. It's about understanding how your
model performs, its strengths, its weaknesses, and whether it actually
meets the business need. Evaluation metrics are crucial for documenting performance,
comparing different models or different versions of the same model,
tracking them over time and production, and importantly for detecting
(24:45):
that model drift we talked about earlier, when performance degrades
metrics tell you it's time to retrain or investigate.
Speaker 1 (24:51):
So it's about ongoing quality control too.
Speaker 2 (24:54):
Definitely for classification models, the ones predicting categories like fraud
not fraud, or spam not spam. The confusion matrix is fundamental.
It's a simple table that breaks down predictions versus actual outcomes.
You get four key numbers. True positives TP correctly predicted positive,
said fraud, was fraud. True negative PN correctly predicted negatives
not fraud, wasn't fraud. False positives FP incorrectly predicted positive,
(25:15):
said fraud, but wasn't. This is a Type I error
of false alarm false negatives. FN incorrectly predicted negative, said
not fraud, but was fraud. This is a type two
error of missed detection.
Speaker 1 (25:24):
TP tn fp FN. Okay, the four outcomes.
Speaker 2 (25:27):
Right, and from this matrix we derive the most common
classification metrics accuracy TP plus TN divided by the total
overall correctness. But careful. It can be really misleading if
your data set is unbalanced like that ninety nine percent
not fraud. Example, recall or sensitivity tp TP plus fn.
This measures how well the model finds all the positive cases.
(25:50):
High recall is crucial when missing a positive is bad
reg missing a disease diagnosis. Precision or positive predictive value
tp TP plus fp this measures how often the model
is correct when it does predict positive High precision is
key when false alarms are costly, for example, marking important
emails as spam.
Speaker 1 (26:07):
Recall finds them all. Precision avoids false alarms. A trade off.
Speaker 2 (26:11):
Often, yes, there's usually a trade off between precision and recall.
The f one score is the harmonic mean of the two,
providing a single score that balances both. Useful when both
precision and recall are important.
Speaker 1 (26:21):
Okay, and what about those curves you see like ROC.
Speaker 2 (26:23):
Right evaluation curves help visualize that trade off across different
decision thresholds. The Precision Recall PR curve plots precision versus recall.
It's particularly useful for imbalanced data sets as it focuses
directly on the performance on the minority class. The ROC
curve receiver operating characteristic plots the true positive rate, which
is just recall, against the false positive rate FP FP
(26:46):
plus TN. It's commonly used for more balanced data sets.
The area under the curve AUC summarizes the curve into
a single.
Speaker 1 (26:52):
Number PR curve for imbalance, ROC for balance. Good tip.
What about regression models? The ones predicting numbers.
Speaker 2 (26:58):
Different metrics there, since we're not dealing with classes. Common
ones include MAE mean absolute error, the average of the
absolute differences between predictions and actual values. Simple intuitive units
MS mean squared error, the average of the squared differences.
This penalizes larger errors much more heavily than smaller ones.
(27:19):
RMS root means squared error the square root of MS.
This brings the metric back into the same units as
your target variable, making it easier to interpret while still
penalizing large errors. RMSE is probably the most common regression metric,
and MAP mean absolute percentage error calculates the error as
an average percentage of the actual values. Very intuitive for
things like sales forecasting.
Speaker 1 (27:40):
Okay, MA rmsc AMAPE for regression. That's a lot of metrics.
If you know, if you're listening and looking at these,
what's maybe one piece of advice you'd give about picking
the right metric for your specific project.
Speaker 2 (27:52):
That's a great question. The single most important thing is
to deeply understand your business goal and the cost of
different types of errors. Don't just default to accuracy because
it sounds good. Ask yourself what's worse a false positive
or a false negative. In medical diagnosis, missing a disease,
a false negative could be catastrophic, so you'd optimize for recall.
(28:13):
In filtering spam, marking a crucial email as spam, a
false positive is highly annoying, so you'd prioritize precision. The
context dictates the metric, Always tie it back to the
real world impact.
Speaker 1 (28:25):
Connect the metric to the business impact excellent advice.
Speaker 2 (28:28):
So once we can measure our models effectively using these metrics,
the next logical step is optimization, specifically hyperparameter tuning. Remember
those knobs that control the learning process. Finding the best
combination of those settings for your specific data is crucial.
The goal isn't just to get a model that performs
well in the data it was trained on. It's to
get a model that generalizes well to new unseen data.
(28:51):
We want to minimize both bias oversimplification and variance overfitting.
Speaker 1 (28:55):
Finding that sweet spot between too simple and too complex.
Speaker 2 (28:58):
Exactly there's several techniques for this hyper parameter search. Grid
search is the most basic. You define a grid of
possible values for each hyper parameter, and it literally tests
every single combination very thorough but can be incredibly slow
and computationally expensive, especially if you have many hyper parameters
or wide ranges. A more efficient approach is random search.
(29:21):
Instead of trying every combination, it randomly samples combinations from
your defined search space. Surprisingly, it often finds very good
or even optimal hyper parameters, much faster than grid search.
Speaker 1 (29:32):
Randomly trying things can be faster.
Speaker 2 (29:34):
Often yes, because not all hyper parameters are equally important.
Random search explores the space more broadly quicker. For even
more intelligence, there's Bayesian optimization. This method learns from past evaluations.
It builds a probability model of how hyper parameters relate
to performance, and uses it to intelligently choose the next
set of hyper parameters to try. Focusing on promising regions
(29:55):
of the search space. It can converge on optimal settings
much much faster, especially for complex models.
Speaker 1 (30:01):
Subaesian learns as it goes.
Speaker 2 (30:02):
Precisely, it's a smarter search.
Speaker 1 (30:04):
The core idea here, then, is that tuning isn't just
a one shot deal. It's an empirical process. You combine
these smart search strategies with solid evaluation techniques, often using
things like cross validation to build robust models, Models that
don't just memorize the training data, but actually perform well
out there in the real world.
Speaker 2 (30:24):
That's the name of the game generalization.
Speaker 1 (30:26):
Okay, bringing this all together. Now we've talked about the
life cycle, data algorithms, evaluation optimization. The guide clearly points
to Amazon stage Maker as the central workbench, the main
hub for doing all this on AWS.
Speaker 2 (30:38):
Yes, sage Maker is designed to be that integrated environment
for the entire mL workflow. It's a fully managed service
aiming to simplify each step. It provides notebook instances. These
are basically managed Jupiter notebooks running on EC two instances.
They're great for data exploration, cleaning, preprocessing, and generally orchestrating
your mL pipeline. For the heavy listing of training, stage
(31:00):
Maker provides dedicated, optimized training instances. You choose the instance
type based on your needs, submit your training code, and
Stagemaker handles provisioning, execution, and tearing down the resources. And
once your model is trained, sage Maker offers endpoint instances
for deploying it and getting real time predictions via a
simple API call. It handles scaling and availability for you.
Speaker 1 (31:21):
Notebooks for exploring training, instances for building endpoints for predicting.
Speaker 2 (31:26):
That's a good summary, and sage Maker also has managed
services specifically for hyper parameter tuning jobs. You de find
your hyper parameters ranges, the metric you want to optimize,
and SageMaker automatically runs the search using strategies like basing, optimization,
random search, or grid search, keeping track of the best performing.
Speaker 1 (31:41):
Job automating that tuning process we just discussed.
Speaker 2 (31:44):
Exactly when you choose instance types for these sage Maker components,
they're all based on EC two instances, often with mL prefixes.
You need to consider your workload. There are general purpose
M family, Compute optimized CE family, Memory optimized OUR family,
and portantly GPU enabled instances like the P and G families.
GPUs are essential for accelerating deep learning training, which involves
(32:08):
massive matrix calculations. The choice really depends on your data size,
algorithm complexity, budget, and how fast you need.
Speaker 1 (32:15):
Results matching the hardware to the mL task. What about security?
Keeping notebooks and data private.
Speaker 2 (32:20):
Security is built in you can launch sage Maker components
like notebook instances or training jobs within your own private
VPC virtual private cloud. This gives you fine grain control
over network access. You can restrict internet access, connect securely
to your on premises data sources, use security groups and
network acls. SageMaker also supports network isolation for training and
(32:41):
inference containers, preventing them from making unauthorized outbound network calls.
Speaker 1 (32:45):
So you can lock it down pretty tightly.
Speaker 2 (32:47):
Absolutely and while sage Maker provides that integrated environment, you
can also orchestrate mL workflows using other AWS services, sometimes
in combination. AWS lamb to functions, ervalless Event driven compute
functions are great for automating parts of the pipeline. For example,
an S three upload event could trigger a lamb to
function to do some initial data preprocessing or validation. For
(33:10):
more complex multi step workflows, AWS step functions is fantastic.
It lets you define your workflow as a visual state machine.
You can sequence and coordinate calls to Lambda functions, glue jobs, SageMaker,
training jobs, manual approval steps, pretty much any AWS service.
It's great for managing long running distributed processes with built
in error handling.
Speaker 1 (33:30):
And retries step functions for orchestrating the whole flow.
Speaker 2 (33:33):
Yeah, and if we connect this back to that bigger picture,
sage Maker, often combined with services like Glue, Lambda and
step functions, really aims to provide a managed, scalable, and
secure environment for that entire Cristium life cycle. We started
with from understanding the business need and preparing data in
notebooks all the way through training, tuning, deploying, and monitoring
(33:54):
models in production. It's designed to be end to end.
Speaker 1 (33:57):
And there you have it. Wow, what an insightful deep
dive that was. We started with the Very Foundation's AIMLDL.
We explored that critical CRISPM life cycle. We understood the
frankly immense importance of data preparation and the variety of
AWS storage options. Then we delved into the specific AWSAI
application services, those ready made tools, and also the powerful
(34:20):
built in algorithms within sage Maker. And finally we saw
how crucial evaluation and optimization are and how sage Maker
and other services help operationalize it. All You've just gained,
I think, a really valuable shortcut to being well informed
about the whole landscape of machine learning on AWS. Extracting
hopefully incredible value from what's normally a pretty dense technical guide.
Speaker 2 (34:40):
And maybe this raises an important final question for you,
the listener, to consider. Given the power you've just heard
about in these integrated AWS services, from intelligent data processing
and diverse storage to that comprehensive suite of AI and
mL tools, how might you reimagine a current data analysis task,
or maybe an automation challenge in your own work? How
(35:01):
could you potentially transform it into an intelligent, scalable mL
solution using some of these capabilities.
Speaker 1 (35:08):
That's a great thought to leave everyone with. How can
you apply this? Thanks for diving deep with us today.
Until next time, keep exploring, keep learning, and stay curious.
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
My Favorite Murder with Karen Kilgariff and Georgia Hardstark
My Favorite Murder is a true crime comedy podcast hosted by Karen Kilgariff and Georgia Hardstark. Each week, Karen and Georgia share compelling true crimes and hometown stories from friends and listeners. Since MFM launched in January of 2016, Karen and Georgia have shared their lifelong interest in true crime and have covered stories of infamous serial killers like the Night Stalker, mysterious cold cases, captivating cults, incredible survivor stories and important events from history like the Tulsa race massacre of 1921. My Favorite Murder is part of the Exactly Right podcast network that provides a platform for bold, creative voices to bring to life provocative, entertaining and relatable stories for audiences everywhere. The Exactly Right roster of podcasts covers a variety of topics including historic true crime, comedic interviews and news, science, pop culture and more. Podcasts on the network include Buried Bones with Kate Winkler Dawson and Paul Holes, That's Messed Up: An SVU Podcast, This Podcast Will Kill You, Bananas and more.