Page 137 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

126 CHAPTER 6 Evolving and Spiking Connectionist Systems

Learning in the SNN is performed in two stages:
• Unsupervised training, where spatiotemporal data is entered into relevant areas
of the SNNcube over time. Unsupervised learning is performed to modify the
initially set connection weights. The SNNcuber will learn to activate the same
groups of spiking neurons when similar input stimuli are presented, also known
as a polychronization effect [96].
• Evolving supervised training of the spiking neurons in the output classification
module, where the same data that was used for unsupervised training is now
propagated again through the trained SNN and the output neurons are trained to
classify the spatiotemporal spiking pattern of the SNNcube into predefined
classes (or output spike sequences). As a special case, all neurons from the
SNNcube are connected to every output neuron. Feedback connections from
output neurons to neurons in the SNN can be created for reinforcement learning.
Different SNN methods can be used to learn and classify spiking patterns from
the SNNcube, including the deSNN [99] and SPAN models [116]. The latter is
suitable for generating motor control spike trains in response to certain patterns
of activity of the SNNr.
Memory in the NeuCube architecture is represented as a combination of the three
types of memory described below, which are mutually interacting:

• Short-term memory, represented as changes of the PSP and temporary changes of
synaptic efficacy;
• Long-term memory, represented as a stable establishment of synaptic efficacyd
long-term potentiation (LTP) and long-term depression (LTD);
• Genetic memory, represented as a genetic code.

In NeuCube, similar activation patterns (called “polychronous waves”) can
be generated in the SNNcube with recurrent connections to represent short-term
memory. When using STDP learning, connection weights change to form LTP or
LTD, which constitute long-term memory.
Results of the use of the NeuCube suggest that the NeuCube architecture can be
explored for learning long (spatio-) temporal patterns and to be used as associative
memory. Once data is learned, the SNNcube retains the connections as a long-term
memory. Since the SNNcube learns functional pathways of spiking activities repre-
sented as structural pathways of connections, when only a small initial part of input
data is entered, the SNNcube will “synfire” and “chain-fire” learned connection path-
ways to reproduce learned functional pathways. Thus, a NeuCube can be used as an
associative memory and as a predictive system with a wide scope of applications.
Using the NeuCube computational platform, application systems can be developed
forlearning,classification,regression,ordataanalysisoftemporal-orspatio/spectrotem-
poraldata.Thefollowingstepsneedtofollowedasadesignandimplementationprocess:
1. Input data transformation into spike sequences;
2. Mapping input variables into spiking neurons;