Page 133 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

122 CHAPTER 6 Evolving and Spiking Connectionist Systems

Based on the ECOS principles, an evolving spiking neural network architecture
(eSNN) was proposed in [37]. It was initially designed as a visual pattern recognition
system. The first eSNNs were based on Thorpe’s neural model [97], in which the
importance of early spikes (after the onset of a certain stimulus) is boosted, called
rank-order coding and learning. Synaptic plasticity is employed by a fast supervised
one-pass learning algorithm. Output neurons evolve in an incremental, online mode,
to capture new data samples. These nodes can merge based on similarity. The eSNN
models use spike information representation, spiking neuron models, and spike
learning and encoding rules, and the structure is evolving to capture spatiotemporal
relationship from data.
Different eSNN models are developed, including:
• Reservoir-based eSNN for spatio- and spectrotemporal pattern recognition
(Fig. 6.7) [98];
• Dynamic eSNN (deSNN) [99]da model that uses both rank-order (RO) and
spike-time dependent plasticity (STDP) learning rules [100] to account for
spatiotemporal data. The training algorithm for deSNN is shown in Fig. 6.8.
Extracting fuzzy rules from an eSNN would make the eSNN not only efficient
learning models, but also knowledge-based models. A method was proposed
[101] and illustrated in Fig. 6.9A and B. Based on the connection weights
(W) between the receptive field layer (L1) and the class output neuron layer (L2),
the following fuzzy rules can be extracted:

IF (input variable v is SMALL) THEN class C i ;
IF(v is LARGE) THEN class C j

FIGURE 6.7
A reservoir-based eSNN for spatiotemporal data classification. Output nodes evolve over
time for each class through supervised learning from input data.