TouchNet preview at Numenta

30
TouchNet preview by Jason Toy

Transcript of TouchNet preview at Numenta

Page 1: TouchNet preview at Numenta

TouchNet preview

by Jason Toy

Page 2: TouchNet preview at Numenta

TouchNet

• A touch simulator and dataset of touchable objects to teach AIs how to interact with their environments via motor-sensory neurons

Page 3: TouchNet preview at Numenta

history

• idea around 2015

• started coding September 2017

• initial release 2017 Q4

Page 4: TouchNet preview at Numenta

why TouchNet?

Page 5: TouchNet preview at Numenta

why TouchNet?

• On Intelligence

• future predicting machines

• neuroscience based principles

Page 6: TouchNet preview at Numenta

why TouchNet?

• Helen Keller

• intelligence through fingers

• consciousness “started” from I

Page 7: TouchNet preview at Numenta

Helen Keller• “Every object is associated in my mind with tactual

qualities which, combined in countless ways, give me a sense of power, of beauty.”

• “..with the fingers I see only a very small portion of a surface, and that I must pass my hand continually over it before my touch grasps the whole. It is still more important, however, to remember that my imagination is not tethered to certain points, locations, and distances. It puts all the parts together simultaneously as if it saw or knew instead of feeling them.”

Page 8: TouchNet preview at Numenta

where is AI now?• images, CNNs, LSTMs, GANs

• most business focus on static models; some work in RL

• where is time?

• motor sensory neurons?

• where is the agent compared to the world?

Page 9: TouchNet preview at Numenta

- John Locke, philosophy of mind

“An unknown, but potentially large, fraction of animal and human intelligence is a direct

consequence of the perceptual and physical richness of our environment, and is unlikely to

arise without it.”

Page 10: TouchNet preview at Numenta

TouchNet consists of 4 parts

• dataset of 3D objects to touch and interact with

• an evolving simulation

• a simple extendable API

• benchmarks to measure and track progress

Page 11: TouchNet preview at Numenta

Dataset

Page 12: TouchNet preview at Numenta

Dataset

• procedurally generated 3D objects

• common objects: cups, plates, bowls, etc

• starting with tool objects

• cifar100 …. imagenet

Page 13: TouchNet preview at Numenta

Simulator

Page 14: TouchNet preview at Numenta

Simulators and physics

• need a real physics engine

• needs to be fully programmable & modifiable

• tested all of them: MuJoCo, DeepMind Lab, Roboschool, project Malmo, VizDoom,Gazebo, bullet , Processing

Page 15: TouchNet preview at Numenta

Simulators and physics• bullet and pybullet are a secret weapon!

• robust physics engine used for many applications

• super easy to install

• movies: Shrek, Hancock,Sherlock Holmes….

• games: Toy Story 3, Grand Theft Auto 4,Hot Wheels…

Page 16: TouchNet preview at Numenta

pybullet• rigid and soft body simulations

• collision detection

• works with unity, blender,CUDA,OpenCL,…..

• developed by Erwin Coumans & community since ~2003

• now at Google and used for RL & Robotics

• http://pybullet.org

Page 17: TouchNet preview at Numenta

What we want to simulate

Page 18: TouchNet preview at Numenta

What we have

Page 19: TouchNet preview at Numenta

simulation demo

Page 20: TouchNet preview at Numenta

coevolving simulator & agent

• index finger with merkel touch sensors

• all sensor types in finger

• all fingers

• full arm

• full body?

Page 21: TouchNet preview at Numenta

API

• simple to get started and use

• easy to install

• modifiable and customizable

Page 22: TouchNet preview at Numenta

reinforcement learning api

• based off OpenAI’s gym and universe API

• can plug in almost any code for reinforcement learning

Page 23: TouchNet preview at Numenta

OpenAI’s Universe API

Page 24: TouchNet preview at Numenta

Same API

Page 25: TouchNet preview at Numenta

Benchmarks

Page 26: TouchNet preview at Numenta

“Blind” objection classification

• goal: classify the object

• CNN and actor critic DNN models trained concurrently

• pytorch and tensorflow versions

• 4% accuracy, please beat mine!

Page 27: TouchNet preview at Numenta

“blind” classification code walk through

Page 28: TouchNet preview at Numenta

Benchmarks

• HTM version?

• more benchmarks in the future e.g:

• moving objects, differentiating textures,…..

Page 29: TouchNet preview at Numenta

Help!

• very ambitious project, too big for me

• 100% open source

• I do a little of everything ( suck at everything? )

• looking for contributors and experts in physics,math, neuroscience, ml

Page 30: TouchNet preview at Numenta

Jason Toy [email protected]

twitter.com/jtoy

Thank you!

Questions and Feedback

http://github.com/jtoy/touchnet