Page 139 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

128 CHAPTER 6 Evolving and Spiking Connectionist Systems

structure consists of spatially allocated spiking neurons, where the location
of the neurons maps a spatial template of the problem space (e.g., brain
template; geographic locations, etc.) if such information exists.
b. The input neurons are spatially allocatedinthisspace to mapthe locationof
the input variables in the problem space. For temporal data for which
spatial information of the input variables does not exist, the variables are
mapped in the structure based on their temporal correlationdthe more
similar temporal variables are, the closer are the neurons they are mapped
into [133].
c. The connections in the SNN are initialized using a smalleworld connectivity
algorithm [6].
2. Encoding of input data:
Input data is encoded into spike sequences reflecting on the temporal changes in
the data using some of the encoding algorithms, for example, Ref. [120].
3. Unsupervised learning in the SNN model:
a. Unsupervised time-dependent learning is applied in the SNN model on the
spike encoded input data. Different spike timeedependent learning rules can
be used. The learning process changes connection weights between individual
neurons based on the timing of their spiking activity.
b. Through learning individual connections over time, whole areas (clusters)
of spiking neurons that correspond to input variables, connect between
each other, forming structural patterns of connectivity of many consec-
utive clusters in a flexible way. The length of the temporal data and
therefore the patterns learned in the SNN model, are theoretically
unlimited.
4. Obtaining dynamic, functional patterns from the SNN model as deep knowledge
representation [6]:
a. A functional, dynamic pattern is revealed as a sequence of spiking activity of
clusters of neurons in the SNN model that represent active functional areas
of the modeled process. Such patterns are defined by the learned structural
patterns of connections.
b. When same or similar input data is presented to a trained SNN model, the
functional patterns are revealed as neuronal activity is propagated through
the connectionist patterns. The length of the obtained functional patterns is
theoretically unlimited.

4.2.1 Supervised Learning for Classification of Learned Patterns in a
SNN Model
When a SNN model is trained in an unsupervised mode on different temporal data,
representing different classes, the SNN model learns different structural and func-
tional patterns. When same data is propagated again through this SNN model, a
classifier can be trained using the known labels, to learn to classify new input
data that activate similar learned patterns in the SNN model.