1. Spatial and Temporal Encoding for a PSN
Student: Cameron Johnson, Department of Electrical and
Computer Engineering
Faculty Advisor: Dr. G.K. Venayagamoorthy, Department of Electrical and
Computer Engineering
RESULTS
The gray-code-based spatial encoding method produced distinctly different numbers and combinations of neurons spiking in response for seven different inputs
OBJECTIVES
Encode real-valued data into a Polychronous Spiking Network (PSN)
Explore the advantages and disadvantages of spatial and temporal encoding methods
Develop means of decoding real-valued data from a PSN after the PSN has performed calculations to generate it
Use a PSN as a function approximator to prove concept
Use a PSN on real-world problems
Robot movement and vision
Power system prediction and control
DISCUSSION
Spatial encoding has yet to yield any decoding information
Shows promise for entering distinct values in and getting distinct responses
Decoding demonstrated by Zhang and Feng via their encoding method
Only manages function reproduction, not calculation
Still promising
The difficulty of encoding and decoding is due to a lack of understanding of how living brains actually use the information they process
Rate coding methods other than Zhang and Feng’s are slow
Spike coding methods typically lose a lot of portential information
BACKGROUND
Current Neural Networks are used for function approximation
Neuroidentification is a form of function approximation
Spiking Neurons model biological neuron behavior more faithfully than other modern neural models
Information is carried and processed by the pattern of spikes which make up the neural network
Translating real data into spikes requires a method of encoding
Temporal encoding has been explored many times, including rate coding and spike time coding
Spatial encoding relies on physical relation of input spikes to the neurons to which they’re connected
RESULTS
CONCLUDING REMARKS
Encoding into a PSN enables a very brain-like model to operate on data
Gain the same sort of intuitive function handling that a living brain can
Navigation through the real world
Expert handling of control problems
Possibly more intuitive handling of instructions
Spatial encoding may deal with long-term memory and learned reflexes
Temporal encoding may deal with pattern recognition and short-term memory
A combination of the two is hoped to exploit the capabilities of a PSN to their fullest
APPROACH: SPATIAL ENCODING
Gray-code-based spatial encoding has been demonstrated in an Izhikevich neural network
Encoding mechanism has
two neurons representing one bit
Neither neuron receiving a spike
means there is no input
A “1” is represented by one
neuron in the pair receiving a
spike and the other nothing.
A “0” is represented by the re-
verse.
APPROACH: TEMPORAL ENCODING
Shown here is an encoding method that treats the real value as a poisson rate applied directly to the voltage equation
X. Zhang, G. You, T. Chen, J. Feng. “Maximum Likelihood Decoding of Neuronal Inputs from an Interspike Interval Distribution.” Neural Computation 21, 2009,
Other traditional
methods
Rate coding
Spike density
Spike count
Population
activity
Spike coding
Time to first
spike
Spike phase
FUTURE WORK
Test temporal decoding methods to see how well they differentiate values
Experiment with combined spatial and temporal encoding
Develop a decoding mechanism for PSNs, whether spatial or temporal or a combination
Use a PSN for power system identification
Acknowledgements
This work was supported by the National Science Foundation (NSF) under EFRI award # and by a GAANN Fellowship.
Special Thanks to:
Real-Time Power and Intelligent Systems Lab
Intelligent Systems Center