TRIANGULATING Intelligence Afternoon Oct 7 2020

<p>TRIANGULATING Intelligence Afternoon Oct 7 2020

sponsored by HAI at Stanford;&nbsp; this session Mike Frank and Chris Manning hosts/ moderate.

First speaker is <strong>Sanjeev Arora</strong>:&nbsp; I missed this talk.&nbsp; (Princeton);

<strong>Yejin Choi</strong>:&nbsp; Commonsense Intelligence from AIAI in Washington

are we done solving AI?

bridge the gap:&nbsp; system 1(intuitive/ instinct)&nbsp; to System 2 (raitonal)

Kahneman's&nbsp; THREE cognit systems from 2003

perception; intuition and then reasoning.

perception:&nbsp; object recog

reasoning:&nbsp; solving puzzles, proving theorems...

but what is in Intuition&nbsp; (system 1)

what humans do every waking day.

what happens before and after.

Roger Shepard's "monsters in a tunnel"

two monsters running&nbsp; (rather than standing still on one foot

one is chasing the other&nbsp; (rather than copying)

the chaser is hostile and the chassed is afreaid&nbsp; (even tho faces are identical)

scene from titanic drowning at end...

Scene graph:&nbsp; does not capture... me n flailing in the water;

why is the person holding ontothe bronze stature.

what is happening&gt;

wha tis actually happening...

new resource&nbsp; visualcomet.xyz

1.4 M&nbsp; inferences from visual scenes.

labesl&nbsp; iamge:&nbsp; server&nbsp; brings platter of food to customers.

Tldr: for reasonign

espec for Inutivie (Commonsense) inferences.

precursor works&nbsp; atomic and comet...

X repels Y's attack...

=&gt;&nbsp; X wanted to protect others;&nbsp; X wantys to learn self defense;&nbsp;&nbsp; etc.

Murdering is 4x as common as exhaling&nbsp; (it might learn cuz no images of exhaling)

OpenCyc:&nbsp; Lenat 1995;

Know. of "what"&nbsp; as in Davis and Marcus in 2015;

Know. of why and how&nbsp; (causes and effects)

ATOMIC&nbsp; Sap in 2019.

Visual Commonsense Graphs&nbsp; (Park in 2020)

ATOMIC has&nbsp; 880k triples for AI systems

do not need to store the fact that butterflies are smaller than elephants.

COMET Commonsenses Transformer.

X&nbsp; rides into the sunset on a matorcycle after solving AI.

Gary breaks the world record for most controversial tweet.

Gary stacks kindling and tosses in a match...
give it to GPT2 and spouts out nonsense.

past failures with commonsense study was due to

weak computation, not muchd ata, no crowdsourcing, etc.

ACL 2020 Commonsense Tutorial&nbsp; (SEE THIS!!!)&nbsp; 2nd most popular at ACL.

so,&nbsp; how to get to the moon?

current paradigm of QA datasets.

we need to tach machines concepts more directly;

must be able to generate the answer (not merely do well on multiple choice);

COMeT

want Symbolic&nbsp; but in language&nbsp; (instad of in LOGIC)

we have 2020 version of ATOMIC coming up.

COMeT&nbsp; based on BART but 435 times smaller.&nbsp; only 400M paras.

GPT3 uses 175B params.

GPT3 morality...

running a blender at 5AM isr ude because

you can wake up the neighborhood

posting fakes news if ok if
it helps the GOP agenda even if it hurts the country.

The Enigma of REason&nbsp; by Hugo Mercier and Dan Sperber... her fave book on this.

Reasoing serves the purpose of communication.. but
reasons are used primarly not to guide oneself but to
justify oneself in theeyes of others and
to convince others.

good... important talk.



<strong>AUDE OLIVA&nbsp; </strong>MIT Direc&nbsp; pron ODE&nbsp;&nbsp; a French woman...

she is MIT Direc of MIT IBM Watson AI Lab... she is French.

Incorp Insights from Cog Sci into AI&nbsp; (models)

from her alb at MIT.

place recog&nbsp; (gist);&nbsp; abstraction and action recog;
and analogy:

what is a meanigful moment/ event....

three seconds is just enuf time to size up a scene....

btwn people, objects, and phenomena&nbsp; (wind blowing, object falling, people hugging)

pushing; carring; breaking;

this window of timeevolved as humans faced physical and corporeal constraints (a body moving thru space, gravity, etc)

and in&nbsp; expts&nbsp; across neural networks 3 seconds are enough to learn video repres.

she shows&nbsp; labeling... skiing, boating, blancing, woring, bounding on trampoline, dancing,
kneeling;&nbsp; 3 seconds is enuf.

Game:&nbsp; which actions best match the reference set...

or which is the odd man out&nbsp; among three short movies.

trying to teach abstraction:

what is the common concept that connects these videos.

shows&nbsp; swimmers competing;&nbsp; cycling competing and basketball game.
which is different?

This is Lee... A Oliva in 2022&nbsp; We have so much in common: modeling: semantic relations set abstraction in videos in ECCV.

examples of abstractions:

transfer object;&nbsp; propel an object;
connect things;&nbsp; separate things;

want to bind new experiences into existing conceptual categories.

now, how to map to the brain:

consumes 20 watts;&nbsp; 100 new neurons per day all your life;
thousand sof difft neuron types;&nbsp; she says 100 Tr connect...

MEG&nbsp; time every millisecond;
fMRI&nbsp; space each millimeter

have subject do same task in MEG and in fMRI.

thisis Cichy, Patazid, and Oliva in 2014 in Nat Neurosci and also
Cichy and Oliva in 2020 Neuron.

they fuse&nbsp; MEG and fMRI...

takes 100 msec to do visual prosessing and on thru 200 and 300 tna d400 msec
etc.&nbsp; etc. thru a full second.

can also do this with sound with headphones... animal sounds,
cars, machines, etc. cars.&nbsp; she stops at 150 msec as sound percolates out
of uperior temp gyrus &nbsp;(audit cortex)
lots of regions are engaged;&nbsp; visual areas that rspond to sounds.

thisis from ben Lahner at MIT.

she shows mind blindness&nbsp; (alternating scens with objects mssing)




<strong>JOSH TENENBAUM....</strong>
SCALING THE AI THE HUMAN WAY&nbsp; (at MIT)

he was an UG at MIT with Roger Shepard and&nbsp; David Rumelhart;

Matt B&nbsp; gave the hx:&nbsp;&nbsp; rats in a maze etc.&nbsp; RL is just one past of intelligence.

we humans build models;&nbsp; we are abt imagining,&nbsp; making models, predicting,
making new models of the world;&nbsp;

start with your kids;

shows kid building jenga tower and kids playing in a bathrub.

This is the best and only example we have...

1.5 year old;&nbsp;&nbsp; make a stack of two cups to make a stack of 5 cups;

2-3 year old&nbsp; opens a door for an adult trying to get into closet!!!

The hard problems of common sense...

generalizaiton:&nbsp; a la Chelsea&nbsp; but in vastly different environments with
no re-training...

shows a guy shooting a basketball.&nbsp; who's helping; who's blocking....

now seeing a different scene still tyrng to hit the basket.

you use physics of all your experience.

or&nbsp; old&nbsp; atari basketball games;

hoops;&nbsp; players shooting&nbsp; (despite primitive graphics)

he wants to understand the abstractions...

Oct 1950&nbsp; MIND
Computing Mahcinery and Intelligenc by A M Turing

Rebecca Saxe;&nbsp; Lix Spelke,&nbsp; Alison Gopnik;&nbsp; Laura...&nbsp; Kyo (student of Laura);

child as scientist;&nbsp; child as coder or haker.

Probabilistic programming languages;

Surya gave hx of the field;

Symbolic langues for repr and reasoningwith abst kno.

Bayesian inference for reasoing abt unobs causes from saprse uncertaind ata

and use neural nets for pattern recog and function approx.

The game engine in the head...&nbsp; how do we build this architecture?

v fast,&nbsp; approx program for simualitng grpahics, physics, and planning;

the game engine is in your head to maek predictions...

Pete Battaglia, (part of Matt's team at DeepMind)

he shows blocks and towers on a table top.

you bump the table hard enuf to knock some blcosk on the floor;
more yellow or more red blocks...

just a few sims and just coarse sims to get the answer.

Kelsey Allen&nbsp; (under Jeff Hamrick) at DeepMind&nbsp; par tof Matt's group)

Virtual tools by&nbsp; Allen, Smith and Tenenbaum in PNAS in press.

jsut takes a few shots.

Intuitive physics in 12 month olds

use infant looing:&nbsp;&nbsp; gumball machine&nbsp; wiht&nbsp; red diamonds vs&nbsp; blue gumballs.

part of brain that encodes&nbsp; heaviness of objects depeding on
how the objects move.

collab with Dan Yamins lab....

now how abt&nbsp; neural models of intuitive physics.

Doris TSAO in 2010&nbsp;&nbsp; generating faces

generates&nbsp; 175 face images...

Intutive pysics in robots....

GEN:&nbsp; a high level programming platform for proab inference and geometric reps.

want to do RT&nbsp; probab inference.

Learning common sense:

a program to learn programs...

OMNIGLOT dataset...

a drwaing program.... how to put ink on the page....
Bayesian inference;&nbsp; to make the strokes....

essentially program synthesis.

The child as hacker&nbsp; (make code more awesome)

tuning parameters&nbsp; (optimize)&nbsp; but also
adding functions;&nbsp; debugging;&nbsp; refactoring code;&nbsp; error handling;&nbsp; extract functions.
write librareis;

work of Kevin Ellis:

Common sense machines&nbsp; (new start up)&nbsp; Marcus and Brooks.

Kevin:&nbsp; DreamCoder:&nbsp; librareis of coidng primitives;

symbolic programming primitives;

do LOGO style&nbsp; pen holding and drawing to make circles, spirals&nbsp; etc.
then expand libarary;
bootstrap to more complex patterns.

DreamCode paper is on Archive&nbsp; (from Kevin Ellis) describes this.

what do u start with and how do you learn the rest...

Program synthesis;

1.5 to 3&nbsp; and then how to read....&nbsp; then use language to learn everythng else.

each stage of kids cognit deve.


QA:&nbsp; it is now&nbsp; 1:55 PM:&nbsp;

Chrtis Manning hosts&nbsp; QA (with Surya)

Sanjeev Arora answers:&nbsp; he is at Princeton:&nbsp; gradient based learning is at the
param level;&nbsp; but the functional aspects of the system are qutie different.
and cannot be inferred from low level stuff).

Josh adds:&nbsp; must have inductive bias:

(I exercise at this point:&nbsp; hopefully this will be avaiable online);

&nbsp; thisis on YouTube!!!

2:45PM&nbsp; Josh Ten.&nbsp; talks about intuition:&nbsp; approx proabab inference
(approx. similarity)&nbsp; can't articulate the reasoning or the basis;
moral judgments...fast and unconsc;

*&nbsp; Final Seesion if&nbsp;&nbsp; 2:40PM to 3:10PM

join by Mike Frankand Bill Newsome...

(they join&nbsp; Chris Manning and Surya Ganguli)

Mike Frank:&nbsp; Symbolic Systems program;&nbsp; cute joke from Chris Manning
(how to reconcile in the midst of the ML/CNN era);
philosophy, law,&nbsp; etc.&nbsp; Stfds take on cog sci.
ourintro course: Mind's and machines.

he himself is an alum of this program;&nbsp;

Bill tells about Wu Tsai: Tsai SNI institute:&nbsp; Wu Tsai:&nbsp; Neurosci Instit

Hennesey and Etchemendy&nbsp; created SNI abt 8 years ago.&nbsp; as multidisc
under dead of research... not in anyone school (eg not in school of med etc);

Too big a probelm (the brain) to be solved by any one discipline.

HAI = Human-Centered AI;&nbsp;

Chris Manning:&nbsp; we need more connective tissue to have fruitful collabs.

John Etchemendy... one of the symbolic systems founders!&nbsp; says Mike Frank.

FAce recog in CNNs&nbsp; won big prize at stfd last year&nbsp; (an udnergrad girl) Neeany?

Q to Bill:&nbsp; how to see neurosci and AI interface:&nbsp; embodied in the new bldg;
Scott Lindemann&nbsp; and Surya and Dan Yamins....
how to handle giant datasets.&nbsp; Bob Laughlin:&nbsp; masses of data and you will simply
be 10 years older.&nbsp; hirng joing faculty... build bridges...
students love AI and neurosci... students are ahead of the faculty hires. (says Bill);

Opport to build a more perfect union;&nbsp;&nbsp;&nbsp; how to build the synergies asks Chris to
Mike FRank:&nbsp; Reid Hoffman&nbsp; program&nbsp; (a new program):&nbsp; Dan Yamins leads one team;
a moonshot;&nbsp;

Bill N:&nbsp; new hires to bridge the gap btwn AI and neurosi and interdisc post-doc program;
to deve leadership of tomorrow;&nbsp; must have people who span the boundaries;

SAYCam ... Mike was a dev of it:&nbsp; Mike bkgrnd in development;&nbsp; GPT3&nbsp;&nbsp; 5 orders of
data more than than child;&nbsp; but Mike's 7 year old can do much more.
Chris Manning works in human language&nbsp; (social interaction is important);

Bill N:&nbsp; agents interacting in an enviro:
Chelsea: working with robots;
Dan working with simualted agents;

babies urge to move:&nbsp; struggling to sit up and stand up; affordances;
kids constantly building structures by eg&nbsp; knocking things over;
big connection between perception and action is crucial:
we see what we know.

This is on YouTube!!!&nbsp; and videos of indiv sessions:




</p>

<p>&nbsp;</p>