Transcript To Understand, Survey and Implement Neurodynamic Models By

To Understand, Survey and
Implement Neurodynamic
Models
By
Farhan Tauheed
Asif Tasleem
Project Progress
► Literature
Review
 Temporal Networks
► Specific
Problem for Implementation
 Implications
► Architectural
Plan for Implementation
 Formal definition
Motivation
► Machine
Perception
► Biological aspects of Traditional Neural
Network Models
 Summation neuron
 Non Linear Activation function
► Non
biological aspects
 Static
 Continuous Input
 Back propagation learning algorithm
Temporal Neural Networks
► Biologically




Inspired
Continuous data feed is operated on
Dynamic Model
Long term Memory
Short term Memory
►Tapped
delay line
►Distributed Time Lagged Feed forward NNs
 Different Back Propagation algorithm
Literature Review
► Universal
Myopic Mapping theorem
 Any uniform fading memory mapped behind a
static network can simulate just as well
► Fontine
and Shastri 1993. have demostrated
that certain tasks not having an explicit
temporal aspect can also be processed
advantageously by Temporal Networks
► Thompson(1996) “Completeness of BSS”
Related Problems
► Time
Series Data Prediction
► Blind Signal Separation
► Cocktail Party Problem
► Attention Based Search Optimization
► Visual Pattern Recognition
Blind Signal Separation Implication
► Speech
Recognition (phoneme recognition)
► Multimedia Compression
► MM database sound based retrieval
► Noise Removal
► Audio Analysis and Visualization
► Sonar and Radar
► Cache Hit Algorithms
Problem Decomposition
► Blind
Signal Separation
 General problem
 No knowledge about the constituents
► Cocktail




Party problem (Specific case)
Much restricted
Few sources
Can be many sensors
Source positioning can also be used as a cue
Continued
► Melody
Decomposition (Specific Case)
 Repetition in constituent signals (Cue)
 Signals usually periodic
 Difficulty (Scale invariant)
► Basic
Keyword “DECONVULUTION”
Cocktail Party Problem
► Formal
Problem Description
 Given N signal sensors receiving N convolved signals
made up of ‘d’ original signals such that d<N
 We have to design an adaptive filter that masks each
original signal from the rest
‘
Our Solution
► Assumptions
 Ideal environment .. No noise .. No other
signals other than ‘d’
 We have prior knowledge of number of ‘d’
 Number of sources is known (MIC) we’re
experimenting with two
► Research
being followed
 COMBINING TIME-DELAYED DECORRELATION AND
ICA:TOWARDS SOLVING THE COCKTAIL PARTY PROBLEM
 By Te-Won Lee & Andreas Ziehe
Solution details
► Network







‘
Architecture
Single layer
Feed forward
Feedback (stability issues )
Sigmoid activation function
Learning rule (Maximizing joint entropy)
Frequency domain…FFT 1024 point
Tapped delay lines for short term memory
Example
Current progress
► Things
Done
 Obtained binaural audio files
 Implementation done in MATLAB
► Using
Neural Network toolbox
► FFT function
 Problem in training time due to FFT in training rule.
► Things
TODO
 Implementation complete / Optimize
 Look into oscillatory networks