January 24, 2019 • 24 mins
Dr. Konrad Kording is a professor at University of Pennsylvania, known for his contributions to the fields of motor control, neural data methods, and computational neuroscience. He runs the Kording Lab, or K-Lab, which focused on computational neuroscience early on and now focuses on causality in data science applications. His lab has made an impact across many fields over time, including Bayesian brains, causal effects in human behavior, and uncertainty in the brain.
In this episode, Matt Taylor interviews Dr. Kording over Skype to discuss illusions of perception, how uncertainty might be represented in the brain, and more. A video of their conversation is available on the HTM School YouTube channel.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Konrad (00:00):
When I don't show you your hands, you don't see your
hand for a while, it turns outthat you become uncertain of
where your hand is because yourproprioception has this thing
that's called"drift." Sobasically if I rotate your hand
here and hold it long enough,you will kind of no longer know
exactly where it is.
Matt (00:19):
Welcome back to Numenta On Intelligence podcast and another
Interview with a Neuroscientist.
A little clip for my interviewwith Dr.
Konrad Kording, who I'll betalking to today.
This is going to be part one ofa two-part series, and I'm also
making some video available ofthis conversation that we had
via Skype, so you can find thaton the HTM School YouTube
(00:40):
channel.
There are a couple ofcommunication blips during this
conversation that hopefully yourbrain's going to work around and
understanding what we aresaying.
Otherwise, we will have part twofor you in February, so I hope
you will listen in then.
Here's Dr.
Konrad Kording.
Okay.
Yeah, I think it's because Iturned the recording on.
Konrad (01:02):
Okay.
Hello.
Recording! And it says you arerecording the call.
Great.
Cool.
Well, nice talking with you.
(02:05):
So, what exactly is theobjective?
I always try to see what'sdriving the whole thing.
Matt (02:10):
So my drive is a couple of things.
First, I'm a community managerfor Numenta and I'm trying to
foster this community ofhobbyists, scientists,
engineers.
It's a very diverse community ofpeople, you know.
When you just open something uplike this, you get a lot of
(02:30):
interesting perspectives.
But my main focus is education,and I really am trying to focus
young people.
With this series of interviews,in general, I'd like to try and
engage people to be interestedin how intelligence works and
how the brain is implementing itand things like that because
that's what got me into it,right?
So I'm sort of focusing on thenew generation of AI experts as
(02:54):
they're growing and I reallywant them to learn how
intelligence works, not just theBayesian models necessarily, you
know.
Konrad (03:03):
Well, of course.
Matt (03:06):
And I'm focusing on how my company is trying to understand
the brain and the discoveriesthat we think we're making, and
the theory that the way we thinkit works because it's theory.
(03:26):
It's theory, and we're not doingexperiments, but we're always
looking at the experimental datato make sure that the theory
still works, you know, in thecontext of these new experiments
and stuff coming up.
And the reason I came to you isbecause you're sort of the type
of neuroscientists now that theway you think about the brain
(03:47):
really aligns well with the waywe think about the brain.
Konrad (03:51):
That's great to hear.
I mean, as you know, I come froma Bayesian background.
Right now, I might be a littlemore connectionist than
Bayesian.
Matt (04:04):
Ebbs and flows.
Konrad (04:04):
But you can't get the Bayesian out of thinking.
For me, I see everything througha Bayesian lens even if I'm not
doing Bayesian things anymore.
Matt (04:16):
That's totally different, and that's great because that's
a perspective that I can'tpresent to my audience because I
didn't come from thatperspective.
By trade, I'm a softwareengineer.
I didn't learn any of the mathsinvolved.
I just learned how to buildthings with software, right?
So I came out of this from anengineering standpoint, how do I
build intelligence?
That was my interest, like whatif something is intelligent and
(04:38):
we can figure out how to reverseengineer it and build it?
So that's really the directionwe're coming from and we're
meeting in the middle.
I mean, eventually we're allgoing to meet in the middle and
build something amazing.
Like that's the vision, right?
Konrad (04:49):
You're right.
You know, you should havestarted the interview.
It starts to be fun.
Like people might find thesetopics interesting already.
Matt (04:56):
We may as well throw the sentiment.
It's recording.
Konrad (05:00):
Awesome.
Cool.
Matt (05:03):
So, great to meet you by the way.
And I haven't introduced you tothe audience or anything but
since we're already recording,the interview's probably
started.
Everyone, this is Dr.
Konrad Kording.
I don't have a good intro.
Let me let you introduceyourself.
I find that it's better forpeople anyway, because I always
mess it up.
Konrad (05:21):
Well, I am not sure if I am good at introducing myself.
So I'm a neuroscientist.
I'm a failed experimentalist.
I started my PhD recording fromprimary visual cortex of the
cat, didn't work all that well,so I ended up thinking of how we
(05:41):
can use math and computation tounderstand the brain.
And I looked at the brain frommany different perspectives over
time.
I was Bayesian for a long timeand thinking about neural
computing my entire life, in away.
I'm interested in deep learningand how the brain might be like
deep learning and how it mightbe different.
(06:04):
And beyond neuroscience, I'minterested in what we can
meaningfully say about the worldusing data.
I'm bad at introducing myself inthat way.
I'm just interested in whateveris interesting.
Matt (06:18):
That's a good way to put it.
I mean, I'm the same way.
I just haven't taken the PhDtrack, and a lot of people in
our community are the same.
We're very passionate abouttrying to understand how things
work.
But there's no major– I getasked this a lot in our
community– What should I studyif I want to understand how
(06:41):
intelligence works?
And this next sort of era of AI,what should I study?
Computational neuroscience orstraight up neuroscience or go
straight with math and physics?
And it's hard to say, really.
Konrad (06:50):
But I think the answer is yes.
Matt (06:52):
Yeah, exactly.
Konrad (06:53):
You should probably study all of them at some level.
If you don't understandpsychology, it's hard for you to
get what intelligence is about.
If you don't study math, it'svery hard to make what you want
to say concrete.
If you don't study computation,you can't actually implement it
and lots of problems you onlysee it once you've tried to
(07:15):
build it.
And similarly, methods fromphysics, they just help you
think about the brain in abroader way, I think.
Matt (07:23):
It took me a very long time to finally come to a
somewhat of a basicunderstanding of intelligence,
and even that is very broad andabstract.
Konrad (07:34):
Please share it.
I'm sometimes very unclear abouthow the brain works.
Matt (07:39):
Oh gosh, don't put me on the spot.
I'm interviewing you! So let meflip it on you, because you said
you used to be a Bayesian thatsort of like, maybe you've
changed your view a bit.
One of the things I found thatwas a theme in your writing or
in the papers that you'veproduced is that if you
structure your experiment in theright way, it's amazing how much
(08:02):
data we can get now fromexperiments and it's just
getting more and more.
And because it's so messy,sometimes you can structure your
experiment that you will findwhat you're looking for
sometimes even though it may notbe what you're looking for.
Right?
Maybe you could talk a bit aboutthat.
Konrad (08:18):
Yeah.
I mean, often times inbehavioral experiments, what
you're looking for is some kindof effect, say"Is uncertainty
relevant when you makedecisions?" So I give you a
task, where uncertainty's reallyuseful, like you knowing how
uncertain you are is reallyessential.
So that's one thing you shoulddo if you're very uncertain and
(08:40):
another thing that you should doif you're very certain about the
situation, and everything elsewill be the same.
So if you want only uncertaintymatters and then yes, you use
uncertainty, and we've shownthat in in many experiments.
But if I give you a situationwhere the reward matters and
(09:01):
nothing else, you'll look forthe reward.
If I give you a situation, whereat some level arousal matters or
something, like how much you'reengaged in a task, well then
that will matter.
If I give you a task where saycolor matters, well then color
matters.
So at some level of uncertainty,there was like in the early two
thousands, like 2005, maybe upto 2010.
(09:27):
There were all of a sudden lotsof labs that said, well let's
see if uncertainty is important,and they all found that
uncertainty is important.
But if we had, instead, gottenall excited about color, then we
might have had brain theoriesthat centered around the idea of
color.
So in the end, when these fieldsreally started going, the
(09:52):
question is really what wedidn't know in the end.
Like we now know thatuncertainty matters, sure.
If I asked you how certain areyou that you will get up between
6:00 AM and 9:00 AM tomorrow,you'll give me a pretty precise
answer to that.
If I asked you how certain it isthat I'll get up in that
(10:14):
interval, which you'll have moreuncertainty, but I'm sure you'll
be less precise.
Matt (10:17):
Right.
Konrad (10:17):
And once I give you some information, you'll be better at
that.
So my view is very much based onuncertainty as that center piece
of the way intelligence works,and you can tell a story where
everything you do is aboutuncertainty.
(10:39):
Now you can tell another story,which is everything is about
learning.
This is a story that you cantell just as well, where you
could say, well if you make amistake, if something goes
wrong, you'll change and nexttime you'll be better.
Now that view can equally, in away, explain a lot of
(11:01):
intelligence, including whatuncertainty gives us.
So uncertainty says along withsomething that I know I kind of
store how certain I am about it,which means that if I'm very
uncertain about something, Iwill rely less on the thing than
if I'm very certain aboutsomething.
(11:22):
Now a learning system will dothe same thing because it
figures out that in the caseswhere it was very certain but
ignored the things, it did worsethan the cases where it very
certain and use the things inthe uncertain cases, but it was
very uncertain about things andvery much relied on it.
That was a mistake.
(11:42):
And so therefore, where are weon that continuum?
I'm not sure.
It could be that the brain–we're born as people who are
there to deal with uncertaintybecause uncertainty is so
important, in which case if youwant uncertainties don't enter
the plan with which our brain ismade.
Or alternatively, we could bereally good at learning and we
(12:05):
would never know the difference.
Matt (12:09):
Would you say that uncertainty is sort of one
dimension of an aspect of ourbrain or it applies to all and
lots of other things that arethat are happening there, or?
Konrad (12:18):
Yeah, I think it is an aspect of reasoning.
It's an important aspect ofreasoning.
Like things that you are moreuncertain about are less
important in the way.
And if you have a bunch ofthings that you're somewhat
uncertain about, each of themyou can combine them to become
(12:40):
more certain.
Matt (12:41):
That's problem solving.
Konrad (12:43):
That's problem solving, exactly.
And dealing with uncertainty isan unavoidable aspect of
intelligence, of problemsolving.
Matt (12:50):
Right.
I visualize this a lot in myhead.
It's hard, right?
When I think about ideas andobjects or things, discrete
things that we think about, youcan apply uncertainty to it.
I like to think about thosethings as multidimensional
attractors, like there's somematch in your brain and some
certain neurons that fire whenyou think about a specific
(13:13):
thing, and the uncertainty aboutthat is sort of how messy and
noisy is that attractor, howwell defined is that thing in
your neurons and the connectionsbetween them.
Konrad (13:25):
That is one way how the brain could deal with
uncertainty.
It's not the only way how youcould deal with uncertainty.
Let's say, it's possible thatkind of, let's say we have some
estimation– is that thing behindyou a guitar that I see?
Now, let's say if my video wassomewhat blocked, it could be
that it's a guitar or could alsojust be that it's like a
(13:48):
painting or something behindyou.
So there's two very differentways how the brain could
represent such a thing.
It could either be, if I'muncertain, if let's say the
image quality is low, my brainactivity is very messy.
But alternatively my brainactivity could be absolutely not
(14:11):
messy at all, where I could saythis could be a guitar, exactly
this guitar with probability0.9, or this could be not messy
at all.
I could say guitar, not guitar.
And so at some level, thismessiness of neural code is
(14:34):
something that communicates toother parts of the brain that we
are uncertain about, but it isonly one of many codes.
You could, for example, say thatthere's just a cell that says,
"How uncertain is my visualsystem at the moment?" in which
case, everything that hasuncertainty has no effect
whatsoever.
No messiness involved, ever.
(14:55):
We basically have one cell thatbasically says, very uncertain,
not very uncertain.
Matt (15:03):
Interesting.
So if you get an ambiguoussensory input and you're trying
to do object identification, youcould, once you match it with
the best thing, you could justsnap.
That's it.
You've made the decision, right?
You've sort of applied yourvision of a guitar.
If you've decided that's aguitar, you make the decision,
I'm going to apply my idea ofwhat a guitar is to that object
(15:25):
in space.
Right?
And it's no longer really you'recertain about.
I mean your certainty, at least,has gone way up because you've
made that decision.
Konrad (15:37):
Right.
Let me show you two areas inwhich case these two ways of
thinking about uncertainty feelvery different.
So let's say one case where yousee a guitar and it's very dark,
so there's a small number ofphotons so you can't be sure if
it's the guitar.
(15:59):
And in that case, you can saymaybe the neural activity is
very massive because kind ofstuff comes in and we're unsure
how to interpret it and that'sless, a small number of photons.
In that case, like thisdisorderliness seems like a very
natural way of thinking, butuncertainty.
Let me take you to another case,where it seems like really
(16:23):
weird.
So I studied movement a lot.
So in movement, when I don'tshow you your hands, you don't
see your hand for a while, itturns out that you become
uncertain of where your hand isbecause your proprioception has
this thing that's called"drift."So basically if I rotate your
(16:46):
hand here and hold it longenough, you will kind of no
longer know exactly where it is.
Matt (16:49):
I know exactly what this feels like because I meditate,
and sometimes I'll wake upfeeling like my arms are in a
different place knowing thatthey aren't really there.
Konrad (17:00):
That is so cool!
Matt (17:00):
But having that sensation that they are somewhere
different.
Konrad (17:02):
There's also the, do you know the rubber hand illusion?
Matt (17:06):
No.
What's that?
Konrad (17:06):
Okay.
Rubber hand illusion is a myth.
So they take a rubber arm that'snot yours, they put it next to
your arm, then they hide youractual arm.
Okay, so there's like now adivider.
Think about it like there's adivider, you see the rubber
hand, you don't see your realhand, and then what they do is
(17:28):
they stimulate the rubber handand your real hand with just
some mechanical device at thesame time.
And after they've done it oftenenough, it feels as if this
rubber hand is totally yourhands.
Matt (17:42):
Yeah, that makes sense to me because you're sort of like
projecting your experience thatyou've several times predicting
that you're going to feel thisand since you see it happening,
even though you don't feel ithappening, I'm like, well that's
good enough.
That's good enough.
Konrad (17:59):
But you feel it happen, that's how they trick your brain
into it, by stimulating it atthe same time.
Whenever you see that the rubberhand gets touched, your real
hand gets touched.
And by doing that often enough,they convince your brain that
the rubber hand's totally yourarm.
And then if someone gets out aknife and stabs into the rubber
hand, you will be scared todeath.
Matt (18:22):
Right.
Konrad (18:23):
So that's the rubber hand illusion, but what it shows
is that you have considerableuncertainty about where your
hand is now when you don't seeit.
Now, here's the interestingthing.
If you want to know how muchuncertainty you have about where
your hand this, it's notsomething fuzzy about the visual
(18:44):
input or something.
It is about you memorizing howlong it's been since you last
saw your hands.
Matt (18:50):
Oh right, right.
Konrad (18:51):
Okay.
So in that case, how uncertainyou are of something is not
something that happens from thevisual stimulus.
It's exactly the same as yourstimulus.
It is something that comes fromsomething that–
Matt (19:03):
Temporal sensations over time, right?
Konrad (19:06):
That's right.
It's something that you need tolearn about.
Something you need to integrate.
Matt (19:11):
Your uncertainty should be based on what you're doing at
any moment in the context ofyour actions.
Konrad (19:15):
That's right.
So in one case, uncertaintysomething that's in the image.
In the other case, it'ssomething that you learned over
time.
It is not in the image at all,but it's something that's
instantaneous.
Matt (19:29):
Right.
Konrad (19:30):
So in that second view, the idea that uncertainty is
sort of something that's in thefuzziness of the neural
representation kind of doesn'tmake much sense because the
stimulus exactly the same.
Where is that uncertainty atthat part coming from?
Matt (19:47):
Some other disconnection up here.
Konrad (19:50):
Yeah.
No, this is weird.
Matt (19:52):
Yeah.
I mean you can convince– this issort of the problem with belief.
You know, you can convinceyourself over time if you're
given enough evidence, that'ssomething completely wrong is
true.
You know, and you believe thatuntil you get enough counter
evidence that you can changeyour beliefs.
(20:13):
And if someone convinces youexperimentally, that's your
hand, that's your hand, that'syour hand.
I can see that they're trickingyou.
They're changing your beliefstructure in your brain for that
small period of time at least.
Konrad (20:26):
That's right.
And you cannot, of course, makea Bayesian argument for that.
Like if the rubber hand actuallyis your hand, it is very
unsurprising that whenever therubber hand gets stimulated, you
feel it.
Matt (20:40):
Right.
Konrad (20:40):
If the rubber hand and your hand would be different,
how improbable is it that theyalways get stimulated at the
same time?
In a way, from a statisticalperspective, your brain does the
right thing.
It's very likely that the rubberarm actually is yours, if you
don't know that theexperimentalists actually
(21:00):
designs the stimulation to trickyou into believing it's the
same.
Matt (21:03):
Right?
Yeah.
That is interesting.
So we started off talking about–sometimes you're not able to
find what you're looking for.
Sometimes just looking forsomething you find a correlation
that might not be the correctcorrelation because you're not
looking at it from a moregeneral aspect.
So I think you've written a bitabout how to do these type of
(21:26):
generalization studies.
Is that a way to counter this,in experiments anyway?
Konrad (21:34):
Yes.
In experiments I think inneuroscience, we really need to
stop doing generalizationsstudies much more.
So what do we typically do is wedo one experiment and then we
have one theory that goes withthat experiment.
In a generalization study, I doone experiment, figure out what
the theory is for that, and thenI do a very different experiment
and see if my theory stillworks.
(21:56):
This is something that we almostnever do in neuroscience and
rarely in psychology.
These resulting theories fromit, they are not put to the
test.
If you, for example say, here'sthe brain area, we recorded from
it.
We find that nuance have tuningto orientation.
Therefore, this must be an areathat does orientation
(22:20):
discrimination.
Matt (22:20):
Hubel and Wiesel stuff.
Konrad (22:21):
Hubel and Wiesel stuff.
And a lot of– and it's the samelogic of a lot of what came
afterwards.
That statement without ageneralization study means
almost nothing for all that wecare, but every feature of the
world that we could change,could change the activity of
those neurons.
There was a recent Carandinistudy that basically showed that
(22:43):
locomotor activity is all overthe visual cortex.
So without a generalizationstudy, it's very hard to know
what we have learned in theoryspace.
Matt (22:56):
So how should neuroscientists go about that?
Well, one of the things you dois you collect data from a lot
of different places, right?
You create your own studiesbecause everyone's got so much
data.
So you think that there's anarea that's ripe for other labs
to try and do sort of thesecross generalization studies
using existing data?
Konrad (23:15):
Yeah, I think times are pretty good for us data
parasites.
Matt (23:22):
Yeah, a lot better than five years ago, I'll tell ya.
Konrad (23:24):
And like I very much believe in there being an
ecosystem for ideas.
Like what is the probabilitythat an experimental lab can
look at the data and figure outstatements that make sense about
intelligence?
(23:44):
Like that glue between data andideas, that is something that we
have in the past often delegatedto purely experimental labs, and
intelligence is prettycomplicated, at least as far as
I'm concerned for the moment.
I think we need an ecosystemwhere people can come up with
(24:07):
ideas, find ways to formalizeit, find ways of testing it, and
that used to be very difficult.
And it's not just on theexperimental side, the same
thing on theories side– thatlots of theorists loath actually
getting their hands dirty andshow that their ideas are born
out in data.
Matt (24:24):
Right.
Konrad (24:24):
I think we need those bridges.
Matt (24:28):
Well, I'm totally on board with that cause.
I think that there's so manygood ideas out there and so many
people approaching this problemfrom different places.
Making those bridges is superimportant so that we can all
build something and progressthis area together in the
future.
So that's the end of part one,folks.
You can continue on listening topart two of this interview with
(24:53):
Dr.
Konrad Kording, when it comesout in February.
Look forward to the next NumentaOn Intelligence podcast.
There'll be another Interviewwith a Neuroscientist.
Thanks for listening.
I'm Matt Taylor from Numenta.
When I don't show you your hands, you don't see your
hand for a while, it turns outthat you become uncertain of
where your hand is because yourproprioception has this thing
that's called"drift." Sobasically if I rotate your hand
here and hold it long enough,you will kind of no longer know
exactly where it is.
Matt (00:19):
Welcome back to Numenta On Intelligence podcast and another
Interview with a Neuroscientist.
A little clip for my interviewwith Dr.
Konrad Kording, who I'll betalking to today.
This is going to be part one ofa two-part series, and I'm also
making some video available ofthis conversation that we had
via Skype, so you can find thaton the HTM School YouTube
(00:40):
channel.
There are a couple ofcommunication blips during this
conversation that hopefully yourbrain's going to work around and
understanding what we aresaying.
Otherwise, we will have part twofor you in February, so I hope
you will listen in then.
Here's Dr.
Konrad Kording.
Okay.
Yeah, I think it's because Iturned the recording on.
Konrad (01:02):
Okay.
Hello.
Recording! And it says you arerecording the call.
Great.
Cool.
Well, nice talking with you.
(02:05):
So, what exactly is theobjective?
I always try to see what'sdriving the whole thing.
Matt (02:10):
So my drive is a couple of things.
First, I'm a community managerfor Numenta and I'm trying to
foster this community ofhobbyists, scientists,
engineers.
It's a very diverse community ofpeople, you know.
When you just open something uplike this, you get a lot of
(02:30):
interesting perspectives.
But my main focus is education,and I really am trying to focus
young people.
With this series of interviews,in general, I'd like to try and
engage people to be interestedin how intelligence works and
how the brain is implementing itand things like that because
that's what got me into it,right?
So I'm sort of focusing on thenew generation of AI experts as
(02:54):
they're growing and I reallywant them to learn how
intelligence works, not just theBayesian models necessarily, you
know.
Konrad (03:03):
Well, of course.
Matt (03:06):
And I'm focusing on how my company is trying to understand
the brain and the discoveriesthat we think we're making, and
the theory that the way we thinkit works because it's theory.
(03:26):
It's theory, and we're not doingexperiments, but we're always
looking at the experimental datato make sure that the theory
still works, you know, in thecontext of these new experiments
and stuff coming up.
And the reason I came to you isbecause you're sort of the type
of neuroscientists now that theway you think about the brain
(03:47):
really aligns well with the waywe think about the brain.
Konrad (03:51):
That's great to hear.
I mean, as you know, I come froma Bayesian background.
Right now, I might be a littlemore connectionist than
Bayesian.
Matt (04:04):
Ebbs and flows.
Konrad (04:04):
But you can't get the Bayesian out of thinking.
For me, I see everything througha Bayesian lens even if I'm not
doing Bayesian things anymore.
Matt (04:16):
That's totally different, and that's great because that's
a perspective that I can'tpresent to my audience because I
didn't come from thatperspective.
By trade, I'm a softwareengineer.
I didn't learn any of the mathsinvolved.
I just learned how to buildthings with software, right?
So I came out of this from anengineering standpoint, how do I
build intelligence?
That was my interest, like whatif something is intelligent and
(04:38):
we can figure out how to reverseengineer it and build it?
So that's really the directionwe're coming from and we're
meeting in the middle.
I mean, eventually we're allgoing to meet in the middle and
build something amazing.
Like that's the vision, right?
Konrad (04:49):
You're right.
You know, you should havestarted the interview.
It starts to be fun.
Like people might find thesetopics interesting already.
Matt (04:56):
We may as well throw the sentiment.
It's recording.
Konrad (05:00):
Awesome.
Cool.
Matt (05:03):
So, great to meet you by the way.
And I haven't introduced you tothe audience or anything but
since we're already recording,the interview's probably
started.
Everyone, this is Dr.
Konrad Kording.
I don't have a good intro.
Let me let you introduceyourself.
I find that it's better forpeople anyway, because I always
mess it up.
Konrad (05:21):
Well, I am not sure if I am good at introducing myself.
So I'm a neuroscientist.
I'm a failed experimentalist.
I started my PhD recording fromprimary visual cortex of the
cat, didn't work all that well,so I ended up thinking of how we
(05:41):
can use math and computation tounderstand the brain.
And I looked at the brain frommany different perspectives over
time.
I was Bayesian for a long timeand thinking about neural
computing my entire life, in away.
I'm interested in deep learningand how the brain might be like
deep learning and how it mightbe different.
(06:04):
And beyond neuroscience, I'minterested in what we can
meaningfully say about the worldusing data.
I'm bad at introducing myself inthat way.
I'm just interested in whateveris interesting.
Matt (06:18):
That's a good way to put it.
I mean, I'm the same way.
I just haven't taken the PhDtrack, and a lot of people in
our community are the same.
We're very passionate abouttrying to understand how things
work.
But there's no major– I getasked this a lot in our
community– What should I studyif I want to understand how
(06:41):
intelligence works?
And this next sort of era of AI,what should I study?
Computational neuroscience orstraight up neuroscience or go
straight with math and physics?
And it's hard to say, really.
Konrad (06:50):
But I think the answer is yes.
Matt (06:52):
Yeah, exactly.
Konrad (06:53):
You should probably study all of them at some level.
If you don't understandpsychology, it's hard for you to
get what intelligence is about.
If you don't study math, it'svery hard to make what you want
to say concrete.
If you don't study computation,you can't actually implement it
and lots of problems you onlysee it once you've tried to
(07:15):
build it.
And similarly, methods fromphysics, they just help you
think about the brain in abroader way, I think.
Matt (07:23):
It took me a very long time to finally come to a
somewhat of a basicunderstanding of intelligence,
and even that is very broad andabstract.
Konrad (07:34):
Please share it.
I'm sometimes very unclear abouthow the brain works.
Matt (07:39):
Oh gosh, don't put me on the spot.
I'm interviewing you! So let meflip it on you, because you said
you used to be a Bayesian thatsort of like, maybe you've
changed your view a bit.
One of the things I found thatwas a theme in your writing or
in the papers that you'veproduced is that if you
structure your experiment in theright way, it's amazing how much
(08:02):
data we can get now fromexperiments and it's just
getting more and more.
And because it's so messy,sometimes you can structure your
experiment that you will findwhat you're looking for
sometimes even though it may notbe what you're looking for.
Right?
Maybe you could talk a bit aboutthat.
Konrad (08:18):
Yeah.
I mean, often times inbehavioral experiments, what
you're looking for is some kindof effect, say"Is uncertainty
relevant when you makedecisions?" So I give you a
task, where uncertainty's reallyuseful, like you knowing how
uncertain you are is reallyessential.
So that's one thing you shoulddo if you're very uncertain and
(08:40):
another thing that you should doif you're very certain about the
situation, and everything elsewill be the same.
So if you want only uncertaintymatters and then yes, you use
uncertainty, and we've shownthat in in many experiments.
But if I give you a situationwhere the reward matters and
(09:01):
nothing else, you'll look forthe reward.
If I give you a situation, whereat some level arousal matters or
something, like how much you'reengaged in a task, well then
that will matter.
If I give you a task where saycolor matters, well then color
matters.
So at some level of uncertainty,there was like in the early two
thousands, like 2005, maybe upto 2010.
(09:27):
There were all of a sudden lotsof labs that said, well let's
see if uncertainty is important,and they all found that
uncertainty is important.
But if we had, instead, gottenall excited about color, then we
might have had brain theoriesthat centered around the idea of
color.
So in the end, when these fieldsreally started going, the
(09:52):
question is really what wedidn't know in the end.
Like we now know thatuncertainty matters, sure.
If I asked you how certain areyou that you will get up between
6:00 AM and 9:00 AM tomorrow,you'll give me a pretty precise
answer to that.
If I asked you how certain it isthat I'll get up in that
(10:14):
interval, which you'll have moreuncertainty, but I'm sure you'll
be less precise.
Matt (10:17):
Right.
Konrad (10:17):
And once I give you some information, you'll be better at
that.
So my view is very much based onuncertainty as that center piece
of the way intelligence works,and you can tell a story where
everything you do is aboutuncertainty.
(10:39):
Now you can tell another story,which is everything is about
learning.
This is a story that you cantell just as well, where you
could say, well if you make amistake, if something goes
wrong, you'll change and nexttime you'll be better.
Now that view can equally, in away, explain a lot of
(11:01):
intelligence, including whatuncertainty gives us.
So uncertainty says along withsomething that I know I kind of
store how certain I am about it,which means that if I'm very
uncertain about something, Iwill rely less on the thing than
if I'm very certain aboutsomething.
(11:22):
Now a learning system will dothe same thing because it
figures out that in the caseswhere it was very certain but
ignored the things, it did worsethan the cases where it very
certain and use the things inthe uncertain cases, but it was
very uncertain about things andvery much relied on it.
That was a mistake.
(11:42):
And so therefore, where are weon that continuum?
I'm not sure.
It could be that the brain–we're born as people who are
there to deal with uncertaintybecause uncertainty is so
important, in which case if youwant uncertainties don't enter
the plan with which our brain ismade.
Or alternatively, we could bereally good at learning and we
(12:05):
would never know the difference.
Matt (12:09):
Would you say that uncertainty is sort of one
dimension of an aspect of ourbrain or it applies to all and
lots of other things that arethat are happening there, or?
Konrad (12:18):
Yeah, I think it is an aspect of reasoning.
It's an important aspect ofreasoning.
Like things that you are moreuncertain about are less
important in the way.
And if you have a bunch ofthings that you're somewhat
uncertain about, each of themyou can combine them to become
(12:40):
more certain.
Matt (12:41):
That's problem solving.
Konrad (12:43):
That's problem solving, exactly.
And dealing with uncertainty isan unavoidable aspect of
intelligence, of problemsolving.
Matt (12:50):
Right.
I visualize this a lot in myhead.
It's hard, right?
When I think about ideas andobjects or things, discrete
things that we think about, youcan apply uncertainty to it.
I like to think about thosethings as multidimensional
attractors, like there's somematch in your brain and some
certain neurons that fire whenyou think about a specific
(13:13):
thing, and the uncertainty aboutthat is sort of how messy and
noisy is that attractor, howwell defined is that thing in
your neurons and the connectionsbetween them.
Konrad (13:25):
That is one way how the brain could deal with
uncertainty.
It's not the only way how youcould deal with uncertainty.
Let's say, it's possible thatkind of, let's say we have some
estimation– is that thing behindyou a guitar that I see?
Now, let's say if my video wassomewhat blocked, it could be
that it's a guitar or could alsojust be that it's like a
(13:48):
painting or something behindyou.
So there's two very differentways how the brain could
represent such a thing.
It could either be, if I'muncertain, if let's say the
image quality is low, my brainactivity is very messy.
But alternatively my brainactivity could be absolutely not
(14:11):
messy at all, where I could saythis could be a guitar, exactly
this guitar with probability0.9, or this could be not messy
at all.
I could say guitar, not guitar.
And so at some level, thismessiness of neural code is
(14:34):
something that communicates toother parts of the brain that we
are uncertain about, but it isonly one of many codes.
You could, for example, say thatthere's just a cell that says,
"How uncertain is my visualsystem at the moment?" in which
case, everything that hasuncertainty has no effect
whatsoever.
No messiness involved, ever.
(14:55):
We basically have one cell thatbasically says, very uncertain,
not very uncertain.
Matt (15:03):
Interesting.
So if you get an ambiguoussensory input and you're trying
to do object identification, youcould, once you match it with
the best thing, you could justsnap.
That's it.
You've made the decision, right?
You've sort of applied yourvision of a guitar.
If you've decided that's aguitar, you make the decision,
I'm going to apply my idea ofwhat a guitar is to that object
(15:25):
in space.
Right?
And it's no longer really you'recertain about.
I mean your certainty, at least,has gone way up because you've
made that decision.
Konrad (15:37):
Right.
Let me show you two areas inwhich case these two ways of
thinking about uncertainty feelvery different.
So let's say one case where yousee a guitar and it's very dark,
so there's a small number ofphotons so you can't be sure if
it's the guitar.
(15:59):
And in that case, you can saymaybe the neural activity is
very massive because kind ofstuff comes in and we're unsure
how to interpret it and that'sless, a small number of photons.
In that case, like thisdisorderliness seems like a very
natural way of thinking, butuncertainty.
Let me take you to another case,where it seems like really
(16:23):
weird.
So I studied movement a lot.
So in movement, when I don'tshow you your hands, you don't
see your hand for a while, itturns out that you become
uncertain of where your hand isbecause your proprioception has
this thing that's called"drift."So basically if I rotate your
(16:46):
hand here and hold it longenough, you will kind of no
longer know exactly where it is.
Matt (16:49):
I know exactly what this feels like because I meditate,
and sometimes I'll wake upfeeling like my arms are in a
different place knowing thatthey aren't really there.
Konrad (17:00):
That is so cool!
Matt (17:00):
But having that sensation that they are somewhere
different.
Konrad (17:02):
There's also the, do you know the rubber hand illusion?
Matt (17:06):
No.
What's that?
Konrad (17:06):
Okay.
Rubber hand illusion is a myth.
So they take a rubber arm that'snot yours, they put it next to
your arm, then they hide youractual arm.
Okay, so there's like now adivider.
Think about it like there's adivider, you see the rubber
hand, you don't see your realhand, and then what they do is
(17:28):
they stimulate the rubber handand your real hand with just
some mechanical device at thesame time.
And after they've done it oftenenough, it feels as if this
rubber hand is totally yourhands.
Matt (17:42):
Yeah, that makes sense to me because you're sort of like
projecting your experience thatyou've several times predicting
that you're going to feel thisand since you see it happening,
even though you don't feel ithappening, I'm like, well that's
good enough.
That's good enough.
Konrad (17:59):
But you feel it happen, that's how they trick your brain
into it, by stimulating it atthe same time.
Whenever you see that the rubberhand gets touched, your real
hand gets touched.
And by doing that often enough,they convince your brain that
the rubber hand's totally yourarm.
And then if someone gets out aknife and stabs into the rubber
hand, you will be scared todeath.
Matt (18:22):
Right.
Konrad (18:23):
So that's the rubber hand illusion, but what it shows
is that you have considerableuncertainty about where your
hand is now when you don't seeit.
Now, here's the interestingthing.
If you want to know how muchuncertainty you have about where
your hand this, it's notsomething fuzzy about the visual
(18:44):
input or something.
It is about you memorizing howlong it's been since you last
saw your hands.
Matt (18:50):
Oh right, right.
Konrad (18:51):
Okay.
So in that case, how uncertainyou are of something is not
something that happens from thevisual stimulus.
It's exactly the same as yourstimulus.
It is something that comes fromsomething that–
Matt (19:03):
Temporal sensations over time, right?
Konrad (19:06):
That's right.
It's something that you need tolearn about.
Something you need to integrate.
Matt (19:11):
Your uncertainty should be based on what you're doing at
any moment in the context ofyour actions.
Konrad (19:15):
That's right.
So in one case, uncertaintysomething that's in the image.
In the other case, it'ssomething that you learned over
time.
It is not in the image at all,but it's something that's
instantaneous.
Matt (19:29):
Right.
Konrad (19:30):
So in that second view, the idea that uncertainty is
sort of something that's in thefuzziness of the neural
representation kind of doesn'tmake much sense because the
stimulus exactly the same.
Where is that uncertainty atthat part coming from?
Matt (19:47):
Some other disconnection up here.
Konrad (19:50):
Yeah.
No, this is weird.
Matt (19:52):
Yeah.
I mean you can convince– this issort of the problem with belief.
You know, you can convinceyourself over time if you're
given enough evidence, that'ssomething completely wrong is
true.
You know, and you believe thatuntil you get enough counter
evidence that you can changeyour beliefs.
(20:13):
And if someone convinces youexperimentally, that's your
hand, that's your hand, that'syour hand.
I can see that they're trickingyou.
They're changing your beliefstructure in your brain for that
small period of time at least.
Konrad (20:26):
That's right.
And you cannot, of course, makea Bayesian argument for that.
Like if the rubber hand actuallyis your hand, it is very
unsurprising that whenever therubber hand gets stimulated, you
feel it.
Matt (20:40):
Right.
Konrad (20:40):
If the rubber hand and your hand would be different,
how improbable is it that theyalways get stimulated at the
same time?
In a way, from a statisticalperspective, your brain does the
right thing.
It's very likely that the rubberarm actually is yours, if you
don't know that theexperimentalists actually
(21:00):
designs the stimulation to trickyou into believing it's the
same.
Matt (21:03):
Right?
Yeah.
That is interesting.
So we started off talking about–sometimes you're not able to
find what you're looking for.
Sometimes just looking forsomething you find a correlation
that might not be the correctcorrelation because you're not
looking at it from a moregeneral aspect.
So I think you've written a bitabout how to do these type of
(21:26):
generalization studies.
Is that a way to counter this,in experiments anyway?
Konrad (21:34):
Yes.
In experiments I think inneuroscience, we really need to
stop doing generalizationsstudies much more.
So what do we typically do is wedo one experiment and then we
have one theory that goes withthat experiment.
In a generalization study, I doone experiment, figure out what
the theory is for that, and thenI do a very different experiment
and see if my theory stillworks.
(21:56):
This is something that we almostnever do in neuroscience and
rarely in psychology.
These resulting theories fromit, they are not put to the
test.
If you, for example say, here'sthe brain area, we recorded from
it.
We find that nuance have tuningto orientation.
Therefore, this must be an areathat does orientation
(22:20):
discrimination.
Matt (22:20):
Hubel and Wiesel stuff.
Konrad (22:21):
Hubel and Wiesel stuff.
And a lot of– and it's the samelogic of a lot of what came
afterwards.
That statement without ageneralization study means
almost nothing for all that wecare, but every feature of the
world that we could change,could change the activity of
those neurons.
There was a recent Carandinistudy that basically showed that
(22:43):
locomotor activity is all overthe visual cortex.
So without a generalizationstudy, it's very hard to know
what we have learned in theoryspace.
Matt (22:56):
So how should neuroscientists go about that?
Well, one of the things you dois you collect data from a lot
of different places, right?
You create your own studiesbecause everyone's got so much
data.
So you think that there's anarea that's ripe for other labs
to try and do sort of thesecross generalization studies
using existing data?
Konrad (23:15):
Yeah, I think times are pretty good for us data
parasites.
Matt (23:22):
Yeah, a lot better than five years ago, I'll tell ya.
Konrad (23:24):
And like I very much believe in there being an
ecosystem for ideas.
Like what is the probabilitythat an experimental lab can
look at the data and figure outstatements that make sense about
intelligence?
(23:44):
Like that glue between data andideas, that is something that we
have in the past often delegatedto purely experimental labs, and
intelligence is prettycomplicated, at least as far as
I'm concerned for the moment.
I think we need an ecosystemwhere people can come up with
(24:07):
ideas, find ways to formalizeit, find ways of testing it, and
that used to be very difficult.
And it's not just on theexperimental side, the same
thing on theories side– thatlots of theorists loath actually
getting their hands dirty andshow that their ideas are born
out in data.
Matt (24:24):
Right.
Konrad (24:24):
I think we need those bridges.
Matt (24:28):
Well, I'm totally on board with that cause.
I think that there's so manygood ideas out there and so many
people approaching this problemfrom different places.
Making those bridges is superimportant so that we can all
build something and progressthis area together in the
future.
So that's the end of part one,folks.
You can continue on listening topart two of this interview with
(24:53):
Dr.
Konrad Kording, when it comesout in February.
Look forward to the next NumentaOn Intelligence podcast.
There'll be another Interviewwith a Neuroscientist.
Thanks for listening.
I'm Matt Taylor from Numenta.
Popular Podcasts
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
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.
Intentionally Disturbing
Join me on this podcast as I navigate the murky waters of human behavior, current events, and personal anecdotes through in-depth interviews with incredible people—all served with a generous helping of sarcasm and satire. After years as a forensic and clinical psychologist, I offer a unique interview style and a low tolerance for bullshit, quickly steering conversations toward depth and darkness. I honor the seriousness while also appreciating wit. I’m your guide through the twisted labyrinth of the human psyche, armed with dark humor and biting wit.