TRIANGULATING INTELLIGENCE Morning Oct 7, 2020
<p>TRIANGULATING INTELLIGENCE Symposium, Oct 7, 2020
Thisis the morning session: 9A to 11:30AM on Oct 7.
This is using VIMEO Livestream (not Zoom) This is a Stanford HAI program on Triangulating Intelligence.
BOTVINICK, Matt, DeepMind,
Matt Botvinick;
starts at 9:15AM (I tune in at 9:30AM); stream is interrupted frequentlu
(very choppy; lots of buffering);
he shows two examples.
second example: facila recog:
Neuro to AI to Psych (is the triangualting intelligence symbol for today)
entire objects are selected by attention and in parallel.
Learning Intuitive Physics
Directional inertia is an example... ball rolling on billiard table;
show toddlers; movies that violate expectations;
develop psych work....
object persistence, solidity, unchangeableness; continuity;
new study by Josh Tenenbaum;
data involves videos and questions; what would have happened if...
counterfactuals; had difficulty;
he includes social science in the traignel (to make it a pyramid)
DAN YAMINS switch to YouTube... less buffering... Dan is at Stfd;
VISION... eg seeing face of Marie Curie;
V1 V2 V4 to IT.
step 1: optimize using CNNS... like a 1000 way Imagenet Task.
eg DiCarlo PNAS 2014.
turns to layer neurons into FACE neurons... by far the best model;
4 principles of goal driven modeling:
1) architecture class eg CNNS of adeq depth
= circuit neuroanatomy
2) task/objective = ecolog niche which is multi-way object categoriz.
3) D = dataset = environment
4 L = learning rule; natl selection and synapt plasticity
doenby arch struct and grad. descent.
PROBLEMS recurrence and feedback
too uch labelled ata requeired.
reall neural netowrks are full of feedback...
there is local recurrence and lots of long range connections.
but require stoo much labeled data:
need reciprolcal data units;
ConRNNs are much better at neural dynamics (relative to strictly FF models)
recurrence is used to cut down # of neurons...
2nd problem: SUPERVISION problem.... no way that babies get millons of difft examples.
SAYCam data set from head mounted infanted age 6 to 32 months
from Mike Frank's group.
Autoencoders:
try to use sparse autoencoders;
colorful image colorization (need to know boundaries ofobjects)
Constrastive Embeddings: push similar things together and dissimilar things further apart
in embedded space.
Chengxu Zhuang his work...
Deep contrastive embedding... see his paper.
want to predict firing in the entire ventral stream hierarchy.
SAYCam Dataset; 3 infants aged 6 to 32 months. with
head mounted cams.
goal is poss model of delveop trajectories....
shift gears to Cog Science can help AI (his part II)
like Matt B: Enviro and an agent (baby or robot eg)
percept and then action...
what happens next.
why we learn world models...
learn across long time horizons...
Flesxible neurla reps for physics problems...
principles from Liz Spelke...
eg object permanence... infants learns at perhaps 3 months.
can do 2d bouncing ball;
ignore pixels: megge pixels into objects to make hierarchi graph states.
rigid sphere rollong out of bowl;
knocking over jenga tower.
difft materials: soft plush toys;
NeurIPS 2020 their paper Flexible neural reps etc.
when you have graph struc; then you can do predictions...
can render back the structures onto the world.
figurs out how to bind things together;
attributes are learned by rendering;
PSGNet predictions.
self-supervideed scene understanding...
want goo model of world but also fit it to the brain regions.
WORLD Model:
what happens next, if I do action A.
also want to learn self model... predict self model and
directs model towared fruiful actions (curiosity can arise)
EGO motion... then gets bored; then plays with objects;
bring objects together so you can bang them together...
capturing intutive pPhysical play;
Active World Model Learning with Curiosity... 2020 paper from his was it ICAM?
animate; sweet spot of understandability (simple, captures essence)
his group Stfd Neuro AI Lab....
CHELSEA FINN is next speaker:
How Not to Create a Robot's Mind: asst prof
start story in 2015; shows PR1 puting objects into slots ina box.
Levine, Finn et al.
JML:R 2016 End to End Training.
robot learning to torque objects into (slots (do not put square pegs into round holes)
same algo to screw cap onto bottle;
and use spatula to life fried egg into bowl.
Robot RL Learning... grew out of FINN 2016 work. and also
inspired Atari and AlphaGo at DeepMind.
BUT... big problem.
but it could not generalize... just ONE spatula (and not all)
perhaps his is like training newborn to use a spatula.
Out training distrib is v v narrow and we want it to be v v braod.
(like the real world)
1st notion: can we build in OUR knowledge of physics etc and just build it in.
But we don't know how to integrate that knowledge into the program.
(that was GOFAI... brittle)
infants, toddles; learn the world by interacting with man objects....
we need to train them in a wider world.
1) how can robots acqurie broad/ divers experiences?
she shows PR2 interacting with lots of objexts... towels...
have human guide robot arm thru various motions:
how about novelty seeking methods:
Plan2Explore Sekar, at ICML in 2020.
How can a user help the robot? but w/o human in the loop exploration.
1) show robot interesting sttes, eg opening a drawer;
after 3 moinutes of human guidance then much better auton exploration.
takeaway: a sm amt of human guidance can go a long way!
next...
collect data next have robot predict the future of objects etc.
models capture genl purpose know. abt the world
and use ALL avaialble supervision signal.
Villegas in NeurIPS 2019.... excellent prediction of behavior of towels when pcked up.
Planning to accomp goals
1) consider potential action sequences
2) predict the future of those actions and
3) select the bast one.
Takeaway: broader data leads to broader generalization
(once we fingoue out how to learn from it)
Open RoboNet DB data from 4+ institutions
data from 7 robot platforms
15 million frames
watching vidoes of people.
she has a dozen studensts;
she teaches CS 330.
Panel Discussion: with Surya, Dan and speakers:
what are the key missing pieces (to get to the moon)?
Matt Botvinick begins answer: we need richer models of the world;
(usu pixel level predictions but that's v limited)
brick at window: can predict stuff without tracing every shard of glass.
are those abstractions important (breaking the wind);
Chelsea: have not exited the atmosphere yet ...
must be able to generalize beyond training distributiosn and beyond pixel level models;
Chris Manning: recurrence and generalization: what's the relation...
trying to get to ho meta learning works... are these different scales than recurrence.
Recurrence: not good if it screws up the state/ learning;
somewhere between vision based objects and symbolic GOFAI objects;
physical scene graphs are crucial (prob more impt than recurrence Dan opines);
Dan jokes... we are climbing a tree... it's the evoltionary tree!!
(v sophisticatedjoke) pun to joke abt getting to the moon.
Matt is a big fan of transformers:;;
Dan likes scene graphs;
Chelsea is skeptical of scene graphs... wants lower level know.
watch out for imposing high level know....
Aud Q: how do attention based models better than RNNs?
how do AI models and brain match? how well does the mapping work?
do the structures that perform the best, map the bes tto the brain...
got to work the real world task
how do they do on fMRI models of activation eg comparing GPT3 and fMRI.
we are climbing the evol. tree.
next Q asked by Chris Manning:
Chris paraphrases Chelsea: we need to train robots on much broader data...
But Chris asks... howmuchis that the solution...
most families aheonly one set of colored blocks and one spatula.
flipping a pancake (Chris M's example)... need to learn the task of flipping
not nec with thousand sof difft spatulas.
*
how to teach robots.... transfer policy to new situations (out of sample);
we do not need to do sim to transfer problem... can just use robots in
the real world.
Chris asks Matt: how to do out of distrib (ie how to generalize):
continual learning (how to do that) impt says Matt B.
how to learn something and use it mauch later,
Matt B example: jack in the box....
how to get bio constraints that are more plausicle than backprop;
Yamins answers: want more bio correct principle of learning rules;
biosci: what the hell are learning rules... they totally fail...
Dan too difficult to observe brain (it's a black box) the brain
is the black box... learning rules: totally inscrutable...
cannot observe what is actually happening.
Dan: bio learning is totally trash!!!
(too much of a black box);
Matt B but revol in methods (eg optogen, calcium imaging) maybe we can
take the next step;
AI is going to need to prime the pump of neuroscientists;
Matt B thinks about AI inspired neurosci.
voltage imaging of synapses in HC in rodent during learning;
Surya speculates: whole system comes to Nash equilib. despite
difft regions using difft learning rules.
Chelsea: what should the learning signal be? and then how do you change the params?
We thought reward prediction error was crucial but it's getting more complex.
MattB: we know what the reward function is in games? points in ATARI or
wins in GO.
robots: what counts as a GOOD imitation: answer may be
humans are the ony real source... to judge whether imitation was good.
opines Matt B;
Matt B: inter. in human choice theory;
need robot parents; Chris brings up necessity of embodiment.
Chris M asks abt robonet; how much do we need;
can you graph the concept of water without feeling water????
Scientist in the crib...
How can the baby possibly know that what is impt is to pay attention to what its
parents are doing; (also intrinsic motivation is important... espec impt in
social interaction)
Bill Newsome: what is boredom in a a digital agent;
shift of distrib: not suffic motivated cuz the motiv is too low;
want to shift your actions to something that will tach you something more or
more efficiently.
Matt I'mwrestling with embodiment issue: but how far can you get on language
learning (but language refers to a world) must have access to the real world;
They go into break at 11:30AM.
</p>
<p> </p>