Bio-inspired computation: Clock-free, grid-free, scale-free, and - - PowerPoint PPT Presentation
Bio-inspired computation: Clock-free, grid-free, scale-free, and symbol-free (FA2386-12-1-4050) PI: Janet Wiles (University of Queensland) AFOSR Program Review: Mathematical and Computational Cognition Program Computational and Machine
(FA2386-12-1-4050) PI: Janet Wiles (University of Queensland)
AFOSR Program Review:
Mathematical and Computational Cognition Program Computational and Machine Intelligence Program Robust Decision Making in Human-System Interface Program (Jan 28 – Feb 1, 2013, Washington, DC)
Bio-inspired computation (PI Janet Wiles)
Research Objectives:
for robots which induce their own temporal terms (symbol-free); use transient micro-synchrony (clock- free); compute at multiple scales (scale-free); and navigate using experience graphs (grid-free). DoD Benefits:
computation in natural systems have the potential to provide new approaches to computation with robust and scalable features. Technical Approach:
that use dendritic computation, different interneuron classes and neural architectures inspired by hippocampus. Budget ($453k): YR 1 YR 2 YR 3 YR 4 151 151 151
Project Start Date: awarded 19 Mar 2012 Project End Date: 1 Sept 2015 (42 months) 2
List of Project Goals 1. Develop a system for extracting multi-scale structure from sequences where the elements in each sequence are not pre-specified and can vary in duration. 2. Develop neural systems that use inhibitory interneurons for clock-free coordination at multiple temporal scales. 3. Develop grid-free algorithms for navigation based on integrated sensory-motor coordination. 4. Test algorithms in simulation and on robot sensory streams using the iRat (a robot developed at UQ for research at the intersection of neurorobotics, neuroscience, and embodied cognition).
3
Progress Towards Goals (or New Goals) 9 months into the project: 1. Progress on a prediction paradigm to extract structure from temporal sequences at multiple scales. Tests on Elman’s simple recurrent network badiiguuu letter prediction task show multiscale term prediction 2. Progress developing neural networks with different inhibitory neuron classes that create an asynchronous temporal pipeline in excitatory neurons 3. Development of a learning algorithm that explicitly adjusts dendritic delays to learn temporally extended patterns. 4. Ongoing development of the iRat as a test platform.
4
5
building on simple recurrent networks (SRN)
6
adiiguuubaguuudiidii
Background Extracting grammatical structure from word prediction Finding word boundaries from letter prediction: badiiguuu
COGNITIVE SCIENCE, 14, 179-211 (1990)
Finding Structure in Time
Jeffrey L. Elman
Time underlies many interesting behaviors.
Simple Recurrent Network
{ba | dii | guuu}*
An advantage is SRNs have unbounded temporal histories
context sensitive grammars
anbn aaabbbaabbaaaabbbb a3b3a2b2a4b4…
7
Learning to predict a context-free language: Analysis of dynamics in recurrent hidden units, Boden, Wiles, Tonkes & Blair ICANN 1999
Background
SRN: 2 inputs, 2 hidden units, 2 output units Task: predict the next input
The limitation is that each SRN has a given spatial scale for both elements and combinations
b b b d d d g g g b a d i i g u u u Syntagmatic (elements that can be combined – phonemes, letters, words) Paradigmatic (elements that can be contrasted) SRNs can use an unbounded past, but predict the future one step at a time
The limitation is that each SRN has a given spatial scale for both elements and combinations
b b b d d d g g g b a d i i g u u u Syntagmatic (elements that can be combined – phonemes, letters, words) Paradigmatic (elements that can be contrasted) ba ba ba dii dii dii guuu guuu guuu b a d i i g u u u Syntagmatic (elements that can be combined) Paradigmatic (elements that can be contrasted) SRNs can use an unbounded past, but predict the future one step at a time Need to represent an unbounded future
Prediction at multiple scales: badiiguuu
10 b d g a i u t-5 12%
10% 0% 15% 26% 36%
t-4 13%
0% 34% 12% 20% 21%
t-3 0%
34% 0% 13% 10% 43%
t-2 32%
0% 0% 0% 34% 34%
t-1 0%
0% 0% 32% 34% 34%
d t
100%
b d g a i u t+1 0%
0% 0% 0% 100% 0%
t+2 0%
0% 0% 0% 100% 0%
t+3 34%
34% 32% 0% 0% 0%
t+4 0%
0% 0% 34% 34% 32%
t+5 13%
10% 11% 0% 34% 32%
b d g a i u t-5 6%
17% 17% 13% 15% 32%
t-4 16%
17% 0% 6% 22% 39%
t-3 16%
0% 0% 16% 34% 34%
t-2 0%
50% 0% 16% 17% 17%
t-1 0%
50% 0% 0% 50% 0%
i t
100%
b d g a i u t+1 17%
17% 16% 0% 50% 0%
t+2 17%
17% 16% 17% 17% 16%
t+3 6%
5% 5% 17% 34% 32%
t+4 12%
10% 12% 6% 22% 38%
t+5 12%
13% 14% 12% 15% 34%
ba dii guuu
t-d
33% 33% 33%
dii
t
100%
ba dii guuu t+d 33%
33% 33%
Time-locked predictions (averaged over 3000 letters) Duration predictions
using dendritic computation and delay learning
time time Spiking neuron
Neural networks with different inhibitory neuron classes can create an asynchronous temporal pipeline in excitatory neurons
11
Defined types of cortical interneurons structure space and spike timing in the hippocampus
(Fig. Somogyi and Klausberger, 2005)
Smith, DeMar, Yuan, Hagedorn and Ferguson, (2001) Delay-insensitive gate- level pipelining, Integration, the VLSI Journal, 30(2):103- 131
Clock-free systems
information
circuit
hardware and more complex circuits
speed when needed (no
and mixtures of rhythms through emergent processes Input- ready Output- ready Data Control Reset
Control points
Proces sing
Output- ready Returning to baseline Input- ready
States
networks: A novel approach to sequence learning based on spike delay variance
Characteristics of structured temporal sequences In neural systems, timing is critical at the millisecond level. Real neural systems:
control, sensorimotor coordination and memory tasks
Learning mechanisms: Neural plasticity when mean and variance are both adaptable
and std dev for x, initially x is normally distributed N(0,1).
within .1 of the target mean, else 0
0.1 0.2 0.3 0.4 0.5
5 10 15 20 25
Challenge: The environment changes and the optimal value for x changes radically. How does the mean value of x move from x=0 to x=20?
0.1 0.2 0.3 0.4 0.5
5 10 15 20 25
Mean 0 St dev: 1, 2, 3, 4, 5, 10, 20
0.1 0.2 0.3 0.4 0.5
5 10 15 20 25
Mean: 0, 5, 10, 15, 20 St dev: 1
0.01 0.02 19.9 19.95 20 20.05 20.1
x = [19.9,20.1]
The signal in the target range [19.9,20.1] is virtually nil
0.01 0.02 19.9 19.95 20 20.05 20.1
The signal can be enhanced by raising the variance.
PDF (50 generations)
0.1 0.2 0.3 0.4 0.5
5 10 15 20 25
X Population variation
Adapt both x and std dev
Probability density functions (1+1 EA, 50 generations)
Message Variance affects fitness signals Variance can move (Gaussian) mountains.
A spiking neural network with spike delay variance learning (SVDL)
2-layer feed-forward network plasticity: gaussian synapses winner-take-all output layer
Wright and Wiles (2012)
Gaussian Synapse Model
p is varied to match a given integral
integral)
Wright and Wiles (2012)
Post synaptic release profiles for mean and variance
(fixed integral) With low variance (0.1), all the current is delivered as a single burst at a delay determined by the mean. At high variances (>5), current is slowly released from the synapse over a large period of time, peaking at the delay determined by the mean.
A B
Wright and Wiles (2012)
Spike Delay-Variance Learning (SDVL) Algorithm
The change of mean, Δμ, is determined by:
where: t0 is the time difference between the presynaptic and postsynaptic spike (ms) μ is the mean of the synapse in milliseconds [min 0, max 15] v is the variance of the synapse [min 0.1, max 10] k(v) is the learning accelerator, here k = (v + 0.9)2 ημ is the mean learning rate α1, α2 are constants
The change of variance, Δv, is determined by:
where: ηv is the variance learning rate β1, β2 are constants
Wright and Wiles (2012)
Results
Synaptic parameter traces and neuron spike times during a sequence recognition task: A. Task design: periods where supervised learning is active. B. The input and output spike times, shown as a phase offset from the beginning of each sequence presentation. C. Mean traces for all synapses in the network. D. Variance traces for all synapses in the network. E. Gaussian postsynaptic release profiles at various stages throughout the trial.
A B C D E
Wright and Wiles (2012)
Winner Takes All SDVL Mechanism: implementation notes
Performs best if:
(period < 5-10Hz)
presentation (at least 3-5)
time step, only allowing SDVL to be performed on the output neuron that has the highest membrane potential
instability when solving the membrane ODE’s
Wright and Wiles (2012)
iRat design goals: rat-sized robots for research at the intersection of neurorobotics, embodied cognition and neuroscience
iRat meets Rat
Rat Excellent navigator, flexible, curious, easily frightened Likes to gnaw things iRat (intelligent rat animat technology) PC on wheels Size is the key challenge for the design of a robot that interacts with a rat Ball, Heath, Milford, Wyeth and Wiles (2010) A Navigating Rat Animat, Alife XII
Distributed intelligence via WLAN antenna Energy via Battery (2 hours continuous use) Local Communication via speakers and microphone Mobility via wheels (over 1.5m/s) Vision via webcam Local control via LCD and navpad Brain via x86 PC 1GHz CPU (RoBoard) Avoidance via IR sensors Robot Operating System (ROS) (Windows or Linux)
iRat
Mass 0.6kg Size 170mm long Ball, Heath, Milford, Wyeth and Wiles (2010) A Navigating Rat Animat, Alife XII
Open RatSLAM
Simultaneous localisation and mapping
Overhead view of environment OpenRatSLAM map
OpenRatSLAM: An Open Source Brain-Based Robotic SLAM System Ball, Heath, Wiles, Wyeth, Corke, Milford (2013), in press.
The OpenRatSLAM module structure
RatSLAM: Micro-experiences create a network of sense impressions (views) linked by motor actions (odometry)
Milford and Wyeth, RatSLAM, 2010; Ball et al 2013 open source RatSLAM
RatSLAM: Micro-experiences create a network of sense impressions (views) linked by motor actions (odometry)
Milford and Wyeth, RatSLAM, 2010; Ball et al 2012 open source RatSLAM
iRat strengths
Platform for
space
Ball, Heath, Milford, Wyeth and Wiles (2010) A Navigating Rat Animat, Alife XII monitors it own battery and recharges autonomously
Left iRat: laser scanner and occupancy grid; Right iRat: forward facing camera and bio-inspired topological map.
iRats with different sensors and mapping systems in conversation
(Heath et al, accepted For ICRA 2013)
List of Publications Attributed to the Grant
Neural Networks: A Novel Approach to Sequence Learning Based on Spike Delay Variance” Presented at WCCI 2012 IEEE World Congress on Computational Intelligence, 2012 - Brisbane, Australia
29