Temporal Hypermap Theory and Application

Preview:

DESCRIPTION

A presentation on the Temporal Hypermap Neural Architecture at the Multimodal Workshop held at Trinity College Dublin, April 2007

Citation preview

A Hypermap Model for Multiple Sequence Processing

Abel Nyamapfene

30 April 2007

Research Motivation

I am investigating complex sequence processing

and Multiple Sequence Processing using an

Unsupervised Neural Network processing

paradigm based on the Hypermap Model by Kohonen

What is A Sequence?

• A sequence is defined as a finite set of pattern items:S: s1 – s2 - … - sn

where sj : j = 1, …, n is a component of the sequence and the length of the sequence is n.

Examples: In Language Processing

• Speech utterances• Action sequences• Gestural sequences

In Multimedia Processing • Video sequences • speech

Why Unsupervised Processing?

• Distributed sequence processing prone to catastrophic interference

• Requirement of teaching signal inappropriate for unsupervised processing applications

Issues in Complex Sequence Processing

• Definition:A sequence is complex if it contains repetitions of the samesubsequence like C- O - N -F-R- O - N -T, otherwise it is a simple

sequence

• Research Issue:In complex sequences the correct sequence component can only be

retrievedby knowing components prior to the current one.

How can this be done automatically in an unsupervised neural network

framework?

Issues in Multiple Sequence Processing

• Built for Single Sequence Processing Only:

Most existing sequential neural networks have no inbuilt

mechanism to distinguish between multiple sequences.

• Research Issue:

How can a sequential neural network learn multiple sequences

one after the other without catastrophic interference – the

phenomenon whereby the most recently learned ones erase the

previously learnt sequences.

The Hypermap Model (Kohonen, 1991)

Context domain selected using input context

vector

Best match from within the selected

context domain picked using input

pattern vector

Key Features• a self-organising map

• patterns occur in the context of other patterns.

• the most recent elements prior to an element, or some processed form of them, used as context.

Shortcomings• Can not recall a sequence using time-varying context• Can not handle multiple sequences

Barreto-Araujo Extended Hypermap Model(1)

a(t-1), y(t-1)

a(t), y(t)

LateralWeights

M

FeedforwardWeights W

z-1z-1z-1z-1 z--1

contextSensory stimuli

Barreto-Araujo Extended Hypermap Model(1)

Key Features:

• Lateral weights encode the temporal order of the sequences.

• The context weights encode sequence identity

• Sensor weights encode sequence item value

Successes:• Can recall entire sequences in correct temporal order • Can handle multiple sequence processing (Up to a point)

Shortcomings:• Model can only handle sequences with no repeating elements.

Barreto-Araujo Extended Hypermap Model(2)

z-1z-1z-1z-1

Fixed context

Time-varyingcontext

Sensorimotor stimuli

a(t-1), y(t-1)

a(t), y(t)

LateralWeights M

FeedforwardWeights W

z--1

Barreto-Araujo Extended Hypermap Model(2)

Key Features:

• Time-varying context vector act as element ID within a

sequence

• Fixed context vector to act as sequence identity vector

Successes:• Can handle both complex and multiple sequences

Shortcomings:• No mechanism to identify, anticipate and recall sequences using partial sequence data• Contextual processing of sequences rather limited

The Hypermap Model for Multiple Sequence Processing

Have Modified the Barreto-Araujo Model as Follows:

• Incorporated a short term memory mechanism to

dynamically encode the time-varying context of each sequence

item, making it possible to recall a stored sequence from its

constituent subsequences.

• Incorporated inhibitory links to enable competitive queuing

during context dependent recall of sequences.

Temporal Hypermap Neuron

Dj0 Dj1 Dj2 Dj(d-1)

Dj+10 Dj+11

Dj+20

Dj+1(d-2)

Dj+2(d-3)

Dj+d-10

Dj+d-21

(j-1)th Neuron (j+1)th Neuron

Pattern VectorContext Vector

Threshold unit

Delay units

Hebbian Link

Hebbian Link

InhibitoryLinks

The Competitive Queuing Scheme for Context- Based Recall

Have Modified the Barreto-Araujo Model as Follows:

• Incorporated a short term memory mechanism to

dynamically encode the time-varying context of each sequence

item, making it possible to recall a stored sequence from its

constituent subsequences.

• Incorporated inhibitory links to enable competitive queuing

during context dependent recall of sequences.

The Competitive Queuing Scheme for Context- Based Recall

• Context vector applied to the network and Winner Take All

mechanism activates all the target sequence neurons

• Inhibitory links ensure that only the first neuron is free to fire.

• Next sequence neuron fires on deactivation of first neuron

• Neuron activation and deactivation continues until entire

sequence is retrieved

• Scheme first proposed by Estes [17], and used by Rumelhart

and Norman [18] in their model of how skilled typists generate

transposition errors.

The Short Term Mechanism for Sequence Item Identification and Recall

• Pattern input Winning Neuron applies a pulse to its tapped delay line.

• Each tap on a delay line feeds into a threshold logic unit. • The inputs to each threshold logic unit are the output of the tap

position to which it is connected on its neuron’s delay line as well as all the simultaneously active tap positions on later neurons in the sequence.

• Threshold unit activation levels and output activations dependent on tap position and WTA ensures highest level threshold logic unit wins the competition.

• STM mechanism transforms network neurons into subsequence detectors which fire when associated subsequence entered into network, one item at a time.

Experimental Evaluation

Evaluation Criteria

• Sought to evaluate the network’s ability to store and recall– Complex sequences– Handle multiple sequences with high degree of overlap

• Sought to compare performance with other models usingPublicly available benchmark data set

Experimental Evaluation1: Evaluation Data

No Sequence No Sequence

1 Learning and Memory II 7 Time Series Prediction

2 Intelligent Control II 8Neural SystemsHardware

3 Pattern Recognition II 9 Image Processing

4 Hybrid Systems III 10Applications of NeuralNetworks to Power Systems

5Probabilistic Neural Networks and Radial Basis Functions

11 Supervised Learning

6Artificially Intelligent Neural Networks II

Network correctly recalls sequences through Context and when partial sequences applied to the network

Partial Sequence Recalled Sequence

Learning and Learning and Memory I1

Radial Radial Basis Functions

Pro No CHOICE due to conflict between sequences 5 and 10

Proc Processing

Time Time Series Prediction

Series Series Prediction

Intelligent No CHOICE due to conflict between sequences 2 and 6

Neural Networks and Neural Networks and Radial Basis Functions

cog cognition II

Artificially Artificially Intelligent Neural Networks II

Hybrid Hybrid Systems III

Case Study:Two-Word Child Language

“there cookie” instead of “there is a cookie”

“more juice” instead of “Can I have some more juice”

“baby gone” instead of “The baby has disappeared”

Two-Word Model Assumptions

• Ordering of two-word utterances not random but similar to adult speech word order (Gleitman et al, 1995)

• Two-word stage and one-word stage communicative intentions similar: e.g. naming, pointing out, state ownership, comment on actions etc (Brown, 1973)

• Leading us to the following two word stage Modelling Assumptions:– a multimodal speech environment as in one-word stage– consistent word ordering for each two-word utterance

Two-Word Model Simulation

• Model based on a Temporal Hypermap with: – Tri-modal Neurons with weights for word utterances,

perceptual entities, and conceptual relations

• Data: – 25 two-word child utterances from the Bloom’73 corpus

Perceptual Entity Vector

WordVector

Conceptual Relation Vector

Inhibitory Link

Z-1 Z-1

(j-1)th Neuron

jth

Neuron

(j+1)th Neuron

Threshold Logic Unit

Delay Line Element

Temporal Hypermap Segment

Discussion: Two-Word Model

• Temporal Hypermap encodes utterances as two-word sequences

• Utterance can be recalled by inputting a perceptual entity and conceptual relationship

• Network capable of utterance completion – Entering a unique first word leads to generation of entire two word utterance

Simulating Transition from One-Word to Two-Word Speech

From Saying

“cookie”

“more”

“down”

To Saying:

“there cookie”

“more juice”

“sit down”

One-Word to Two-Word Transition Model Assumptions

• From 18th month to 24th month child language undergoes gradual and steady transition from one-word to two-word speech (Ingram,1981; Flavell 1971)

• Transition is gradual, continuous and has no precise start or end point (Tomasell & Kruger, 1992)

• For the transition we assume for each communicative intention:– Transition probability increases with exposure (training)– Transition is non-reversible transition

Gated Multi-net Simulation of One-Word to Two-Word Transition

• Gated Multinet Architecture comprises: – a modified counterpropagation network for one word

utterances – a Temporal Hypermap for two-word sequences– Exposure-dependent inhibitory links from Temporal

Hypermap units to counterpropagation network to manage transition from one-word to two-word output

• Data: – 15 corresponding pairs of one-word and two-word child

utterances from the Bloom’73 corpus

One-Word to Two-Word Model

Static CPnetwork

TemporalHypermap

Inhibitory Link

Wordvector

Perceptual entityvector

Conceptual Relationship

vector

z-1z-1

Output Two-Word Utterances Plotted Against Number of Training Cycles

0.00

2.00

4.00

6.00

8.00

10.00

12.00

1 3 5 7 9 11 13 15 17 19

No. of Training Cycles

Tw

o W

ord u

tterances

P=0.001P=0.01P=0.1

Future Work:Majority of multimodal models of early child language are static:

Image (input)

Image (output) Label (output)

Label (input)

Plunkett et al model, 1992

perceptual input

Conceptual Relation input

One Word Output

Nyamapfene & Ahmad, 2007

• Spoken words can be viewed as phoneme sequences and/or syllable sequences

• Motherese directed at preverbal infants relies heavily on the exaggerated emphasis of temporal synchrony between gesture/ongoing actions and speech (Messer, 1978; Gogate et al, 2000; Zukow-Goldring, 2006)

But in, reality, the early child language environment is both multimodal and temporal:

So we intend to model early child language as comprising phonological word forms and perceptual inputs as temporal sequences

We Will Make These Modifications

Dj0 Dj1 Dj2

(j-1)th Neuron (j+1)th Neuron

Pattern VectorContext Vector

Threshold unit

Delay units

Hebbian Link

Hebbian Link

To includePattern Items

from other sequences

1

To IncludeFeedback from

Concurrent Sequences

2

Thank YouDiscussion and Questions

??!!

Recommended