Transcript - Backpack

Google Brain
Gagan Khanijau
2011046
Prabhat Ranjan
MT13046
Plan
 Introduction – Artificial Brain
 Google Brain
 Experiment
 Applications
 The Pattern Recognition Theory of Mind
 A Strategy for creating a Mind
 References
Introduction – Artificial Brain
 Aims to reproduce capabilities similar to a human
or an animal brain using the concepts of AI and
machine learning
 Similar research:
 Blue Brain – IBM
 Human Brain Project - École Polytechnique Fédérale de
Lausanne
 IBM 5 in 5, 2012
Google Brain
 Google’s Deep Learning project which later
acquired the name “Google Brain”
 Aimed to mimic some aspects of Human Brain
 Currently, has been successfully trained to
recognize a cat based on 10 million Youtube
images.
 Project was initiated by Andrew Ng and currently
also includes Ray Kurzweil, Jeff Dean, Geoffrey
Hinton in the team
Google Brain (Cont.)
 Based on self taught learning and deep learning
technologies.
 Uses a large scale neural network for performing a
standard image classification test.
 Unsupervised learning.
Experiment
 Connected 16,000 computer processors and let
the network turned loose on the Internet to learn
on its own.
 Feeding the neural network 10 million random
digital images from YouTube videos.
 The unsupervised machine taught itself to
recognize felines, it invented the concept of cat.
Applications
 It is leading to significant advances in areas as
diverse as machine vision and perception, speech
recognition and language translation.
 The project's technology is currently used in
the Android OS’s speech recognition system and
photosearch for Google+.
The Pattern Recognition Theory of Mind
 Neocortex : Responsible for our ability to deal with
pattern of information and to do so in hierarchical
fashion.
 Accounting 80% of total weight of brain.
 Large forehead means larger neocortex.
 Human cortex basically made up of 6 layers.
 Numbered I (the outermost layer) to VI
 The axons emerging from neurons in layers II and III
project to others parts of the neocortex.
 The axons(output connections) from layers V and VI are
connected primarily outside of the neocortex to the
thalamus, brain stem, and spinal cord.
The Pattern Recognition Theory of Mind..
 The neurons in layer IV receive synaptic (input)
connections from neurons that are outside the
neocortex, especially in the thalamus.
 Basic unit of neocortex
 “cortical column(Pattern recognizer)”
 Human neocortex contains :
 300 million pattern recognizer
 Each pattern recognizer consist of 100 neurons
 Redundancy
 There is more than one pattern recognizer for a input.
 It increases the likelihood of successful recognition
 Hierarchy of concepts
A Strategy for creating a Mind
“There are billions of neurons in our brain , but
what are neurons? Just cells. The brain has no
knowledge until connection are made between
neurons. All that we know, all that we are, comes
from the way our neurons are connected”
-Tim Berners-Lee
 Start with building a pattern recognizer that meets
the necessary attributes.
 Next, make so many copies of the recognizer as we
have memory and computational resources to
support.
A Strategy for creating a Mind….
 Each recognizer computes the probability that its
pattern has been recognized.
 Takes into consideration the observed magnitude of each
input.
 Recognizer triggered its stimulated axon if that
computed probability exceeds a threshold.
{Threshold and the parameter that control the
computation of the pattern’s probability are
among the parameter , we will optimize with a
generic algorithm.}
 Recognition of the pattern sends an active signal
up the simulated axon of this pattern recognizer.
A Strategy for creating a Mind….
 This axon is in turn connected to one or more
pattern recognizers at the next higher conceptual
level.
 The pattern recognizers are responsible for
“wiring” themselves to others pattern recognizers
up and down the conceptual hierarchy.
Note: Implementation of “wires” in software
systems is done via virtual links(basically memory
pointers).
 We simply assign new memory locations to a new
pattern recognizer and use memory links for the
connections.
A Strategy for creating a Mind….
 Mathematical Technique :
 For self-organizing hierarchical pattern recognition, we
use HHMM(Hierarchical hidden markov model).
 Accommodate substantial redundancy of each pattern
especially that occurs frequently, we use Linear
programming.
 Education of Brain
 Old learned Brain
 Provide critical thinking module
 Perform continual background search
 Modules for open questions in every descipline
 Give time to evolve.
A Strategy for creating a Mind….
 Finally, Our new brain needs a purpose.
 A purpose is expressed as a series of goals.
 Like IBM Watson’s goal was to respond to “Jeopardy” queries.
 More interestingly, we could give our new brain a
more ambitious goal, such as contributing to a
better world.
 Goals , of course raise a lot of questions:




Better for whom?
Better in what way?
For biological humans?
For all conscious beings?
 Moral educations for our new brain.
References
 http://en.wikipedia.org/wiki/Google_Brain
 http://www.npr.org/2012/06/26/155792609/a-massivegoogle-network-learns-to-identify
 http://newsfeed.time.com/2012/06/27/google-builds-abrain-that-can-search-for-cat-videos/
 http://googleblog.blogspot.in/2012/06/using-large-scalebrain-simulations-for.html
 “How to Create A Mind” by Ray Kurzweil