Page 135 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

124 CHAPTER 6 Evolving and Spiking Connectionist Systems

eSNN have several parameters that need to be optimized for an optimal perfor-
mance. Several successful methods have been proposed for this purpose, among
them are: quantum-inspired evolutionary algorithm (QiEA) [31] and quantum-
inspired particle swarm optimization method (QiPSO) [32].
Quantum inspired optimization methods use the principle of superposition of states
to represent and optimize features (input variables) and parameters of the eSNN [37].
Features and parameters are represented as qubits, which are in a superposition of 1
(selected) with a probability a, and 0 (not selected) with a probability b.When the
model has to be calculated, the quantum bits “collapse” into a value of 1 or 0.

3.2 APPLICATIONS AND IMPLEMENTATIONS OF SNN FOR AI
Numerous applications based on different SNN and more specifically, on eSNN, are
developed, for example,

• Advanced spiking neural network technologies for neurorehabilitation [102];
• Object movement recognition [103];
• Multimodal audio and visual information processing [104];
• Ecological data modeling and prediction of the establishment of invasive species
[105];
• Integrated brain data analysis [106];
• Predictive modeling method and case study on personalized stroke occurrence
prediction [107].
The full advantage of SNN in terms of speed and low computational cost can be
achieved when SNN are implemented in neuromorphic hardware platforms. Contrary
to the traditional von Neumann computational architecture, where memory, control,
and ALU are separated, in neuromorphic systems all these modules are integrated
together as they are integrated in the brain.
To make the implementation of SNN models more efficient, specialized neuro-
morphic hardware are developed, including:

• A hardware model of an integrate-and-fire neuron [108];
• A silicon retina [109];
• INI Zu ¨rich SNN chips [110,111];
• IBM True North [112]. The system enables parallel processing of 1 million
spiking neurons and 1 billion synapses;
• DVS and silicon cochlea (ETH, Zurich) [113];
• Stanford NeuroGrid [114]. The system has 1 million neurons on a board,
63 billion connections, and is realized as hybrid analog/digital circuits;
• SpiNNaker [115]. The system is a general-purpose, scalable, multichip multicore
platform for real-time massively parallel simulations of large-scale SNN.
The neuromorphic platforms are characterized by massive parallelism, high
speed, and low power consumption. For their efficient application, they require
the development of SNN computational models for learning from data.