Transcript Quality – An Inherent Aspect of Agile Software Development

Hierarchical Temporal Memory
“The Classification of Un-preprocessed
Waveforms through the Application of the
Hierarchical Temporal Memory Model”
April 18, 2009
Version 1.0; 04/18/2009
John M. Casarella
Ivan G. Seidenberg School of CSIS, Pace University
Topics to be Discussed

Introduction

Intelligence

Artificial Intelligence

Neuroscience

Connectionism and Classical Neural Nets

Pattern Recognition, Feature Extraction & Signal Processing

Hierarchical Temporal Memory (Memory – Prediction)

Hypothesis

Research

Results
Introduction
"I was proceeding down the road. The trees
on the right were passing me in orderly
fashion at 60 miles per hour. Suddenly one
of them stepped in my path."
John von Neumann providing an
explanation for his automobile accident.
Intelligence

What is Intelligence?





A uniquely human quality?
“means the ability to solve hard problems”
Ability to create memories
Learning, language development, memory
formation (synaptic pattern creation)
The human ability to adapt to a changing
environment or the ability to change our
environment for survival
Alan Turing






“Can machines Think?”
Connectionism
Model digital computer like child’s mind, then
“educate” to obtain “adult”
“Unorganized Machines” : A network of neuronlike Boolean elements randomly connected
together
Proposed machines should be able to ‘learn by
experience’
The Turing Test - constrained and focused
research


Imitate human behavior
Evaluate AI only on the basis of behavioral response
Turing’s unorganized machine
Machine Intelligence

What is Artificial Intelligence?


The Objectives of AI




Create machines to do something which would require
intelligence if done by a human
To solve the problem of how to solve the problem
von Neumann and Shannon


The science and engineering of making intelligent machines,
especially intelligent computer program
Sequential processing vs. parallel
McCarthy (Helped define AI), Minsky (first dedicated AI lab
at MIT) and Zadeh (Fuzzy Logic)
Various Varieties



Expert Systems (rule based, fuzzy, frames)
Genetic Algorithms
Perceptrons (Classical Neural Networks)
Neural Nets

McCulloch and Pitts







Model of Neurons of the brain
Proposed a model of artificial Neurons
Cornerstone of neural computing and neural networks
Boolean nets of simple two-state ‘neurons’
Concept of ‘threshold’
No mechanism for learning
Hebb - Pattern recognition learned through by
changing the strength of the connection between
neurons
Classical Neural Networks

Rosenblatt
Perceptron Model - permitted mathematical analysis of
neural networks
 Based on McCulloch and Pitts
 Linear combiner followed by a hard limiter
 Activation and weight training
 Linear Separation - No XOR

Classical Neural Networks

Minsky and Papert, what where they thinking
Mathematical proof perceptron model of limited usefulness
 Classes of problems which perceptrons could not handle
 Negative impact on funding
 Narrow analysis of the model
 Incapable of learning the XOR - wrong
 Incorrectly postulating that multi-layer perceptrons would
be incapable of the XOR

Classical Neural Networks


The Quiet Years, Grossberg, Kohonen and
Anderson
Hopfield



Rumelhart and McClelland



Introduced non-linearities
ANNs could solve constrained optimization problems
Parallel Distributed Processing
Backpropagation
Interdisciplinary nature of neural net
research
Neuroscience



Structure of the neocortex
Learning, pattern recognition and synapses
Mountcastle

Columnar model of the Neocortex
Learning associated with construction of cell
assemblies related to the formation of pattern
associations


Neuroplasticity
Neuroscience

The Biology





Neocortex >50% of human brain
Locus of: perception, language, planned behavior,
declarative memory, imagination, planning
Extremely flexible/generic
Repetitive structure
The Neocortex




Hierarchy of cortical regions
Region - region connectivity
Cortical - thalamic connectivity
Cortical layers: cell types and connectivity
Neuroscience






Layer 1
Layer 2
Layer 3

A

B
Layer 4
Layer 5

A

B
Layer 6

A

B
Electrocardiogram






Electrocardiogram (ECG) records the
electrical activity of the heart over time
Breakthrough by Willem Einthoven in
1901
Electrodes placed as per a pattern
ECG displays the voltage between pairs
of these placed electrodes
Immediate results
Assigned letters P, Q, R, S and T to the
various deflections
Electrocardiogram
Measurement of the
flow of electrical
current as it moves
across the
conduction pathway
of the heart.
Recorded over time
Represents different
phases of the cardiac
cycle
Electrocardiogram
Hypothesis


Application of the HTM model, once
correctly designed and configured,
will provide a greater success rate in
the classification of complex
waveforms
Absent of pre-processing and
feature extraction, using a visual
process using actual images
Research

Task Description



Create an image dataset of each waveform
group for classification
Determine, through organized experiments,
an optimized HTM
Apply optimized HTM to the classification of
waveforms using images, devoid of any preprocessing or feature extraction
Hierarchical Temporal Memory
Overview


Each node performs similar
algorithm
Each node learns




“Names” of groups passed up



Hierarchy of memory nodes

1) Common spatial patterns
2) Common sequences of spatial
patterns
(use time to form groups of patterns
with a common cause)
- Many to one mapping, bottom to
top
- Stable patterns at top of hierarchy
Modeled as an extension of
Bayesian network with belief
propagation
Creates a hierarchical model (time
and space) of the world
Hierarchical Temporal Memory
Structure of an HTM
network for learning
invariant representations for
the binary images world.
 This network is organized
in 3 levels. Input is fed in at
the bottom level. Nodes are
shown as squares.
 The top level of the
network has one node, the
middle level has 16 nodes
and the bottom level has 64
nodes.
 The input image is of size
32 pixels by 32 pixels.
This image is divided into
adjoining patches of 4 pixels
by 4 pixels as shown.
Each bottom-level node’s
input corresponds to one
such 4x4 patch.

Hierarchical Temporal Memory
The fully learn ed node has 12 quantiz ation centers within its spatial po oler and 4 tempo ral
groups within its tempo ral po oler . The quantiz ation cente rs are shown as 4x4 pix el patc hes.
The tempo ral groups are shown in ter ms of their quantiz ation centers. The input to th e node
is a 4x 4 pix el patch .
Hierarchical Temporal Memory
QuickTime™ and a
decompressor
are needed to see this picture.
This figure illustrates how nodes operate in a hierarchy; we show a two-level network and its associated inputs for three
time steps. This network is constructed for illustrative purposes and is not the result of a real learning process. The outputs
of the nodes are represented using an array of rectangles. The number of rectangles in the array corresponds to the length
of the output vector. Filled rectangles represent ‘1’s and empty rectangles represent ‘0’s.
Hierarchical Temporal Memory
QuickTime™ and a
decompressor
are needed to see this picture.
This input sequence is for an “L” moving to the right. The level-2 node has already learned
one pattern before the beginning of this input sequence. The new input sequence
introduced one additional pattern to the level-2 node.
Hierarchical Temporal Memory




(A) An initial node that has not
started its learning process.
(B) The spatial pooler of the
node is in its learning phase
and has formed 2
quantization enters
(C) the spatial pooler has
finished its learning process
and is in the inference stage.
The temporal pooler is
receiving inputs and learning
the time-adjacency matrix.
(D) shows a fully learned
node where both the spatial
pooler and temporal pooler
have finished their learning
processes
Hierarchical Temporal Memory
Temporal Grouping
QuickTime™ and a
decompressor
are needed to see this picture.
Hidden Markov Model - A
Temporal Grouping
QuickTime™ and a
decompressor
are needed to see this picture.
HMM - 2
Temporal Grouping
QuickTime™ and a
decompressor
are needed to see this picture.
HMM - 3
Experimental Design

HTM Design, Parameters and Structure





Determine the number of hierarchies to be used
Image size in pixels influences the sensor layer
Pixel Image size broken down into primes to determine
layer 1 and layer 2 array configuration
Determine the number of “iterations” of viewed images
at each layer
Small learning and unknown datasets used
Hierarchical Temporal Memory





Memorization of the input patterns
Learning transition probabilities
Temporal Grouping
Degree of membership of input pattern
in temporal group
Belief Propagation
Experimental Design

Waveform Datasets





Individual beats broken down by classification and
grouped (SN, LBBB, RBBB)
Teaching and unknown dataset randomly created
Teaching sets of 50, 90 and 100 images used
Multiple sets created
Teaching vs. Unknown Datasets


Traditional ratios 1:1 or 2:1, teaching to unknown
With HTM model, the ratio was 1:3, teaching to
unknown
ECG Waveform Images - Sinus
ECG Waveform Images
Individual ECG Beat Images
Left Bundle Branch Block
Normal Sinus
Right Bundle Branch Block
All images were sized to 96 x 120 pixels
ECG Series Waveform Images
Left Bundle Branch Block
Right Bundle Branch Block
Sinus
Results – Individual Beat Classification

Learning





Smaller number of teaching images
Diverse images produce greater classification pct
Overtraining not evident, saturation may exist
RAM influences performance
Waveform Datasets




Diversity
Noise : approx. 87 pct w/o inclusion in teaching set
Average > 99 percent classification
Average differentiation of images by class approx. 99 pct
NUPIC Model Results
48 object categories producing 453 training images. Only 99.3 percent of the training
images were correctly classified. (32 x 32 pixels)
Of this “distorted set”, only 65.7 percent were correctly classified within their
categories
Individual Beat Results
Percent Classified by Model
HTM Model Series
Percent Classified
htm_100_17
98.5
htm_100_19
99.2
htm_100_21
98.8
htm_100_23
99.2
htm_100_25
99.2
htm_100_29
99.0
htm_100_33
98.8
htm_100_37
98.8
Results by Dataset
100
99.8
99.6
99.4
99.2
99
98.8
98.6
98.4
98.2
98
p dataset
q dataset
r dataset
s dataset
t dataset
v dataset
x dataset
Results by HTM Model
100
99.8
99.6
99.4
99.2
99
98.8
98.6
98.4
98.2
98
ht m_ 100_ 17
ht m_ 100_ 19
ht m_ 100_ 21
ht m_ 100_ 23
ht m_ 100_ 25
ht m_ 100_ 29
ht m_ 100_ 33
ht m_ 100_ 37
IR Spectra
Sample IR Spectra
Results – IR Spectra
Figure X: Left: IR Spectrum of an Alcohol; Right: IR Spectrum of a Hydrocarbon
Figure X: IR Spectrum of a Nitril e
Classification pct > 99
Gait Waveforms
ALS
Control
With a limited teaching and
unknown set > 98 pct
Huntingtons
References (Short List)







[4] Computational Models of the Neocortex. http://www.cs.brown.edu/people/tld/projects/cortex/
[6] Department of Computer Science, Colorado State University.
http://www.cs.colostate.edu/eeg/?Summary.
[8] George, Dileep and Jaros, Bobby. The HTM Learning Algoritims. Numenta, Inc. March 1, 2007.
[11] Hawkins, Jeff and Dileep, George. Hierarchical Temporal Memory, Concepts, Theory, and Terminology.
Numenta, Inc. 2006.
[12] Hawkins, Jeff. Learn like a Human. http://spectrum.ieee.org/apr07/4982.
[15] Swartz Center for Computational Neuroscience. University of California San Diego.
http://sccn.ucsd.edu/eeglab/downloadtoolbox.html
[16] Turing, A. M. “Computing Machinery and Intelligence”. Mind, New Series, Vol. 59, No. 236. (Oct.,
1950). pp 443 – 460.