Netlab Loligo: New Approaches To Neural Network Simulation

This book describes a programmer's attempt to create a software abstraction that embodies some of the understanding gained in the fields of neurobiology, behavior, cognition, learning, and philosophy of mind. A variety of novel concepts and methods have emerged from the exploration, which are described within.

Constant Learning €" Taken together, the described innovations overcome the obstacles normally encountered in the design of systems capable of continuously learning and adapting within a complex milieu. This is a primary driving goal in the design of Netlab.

No Restrictions On Feedback €" A new learning method, called Influence Learning, is presented that is based on a given neuron using its more influential post-synaptic neurons as role models. Existing neural network models often impose restrictions in the amount, or type, of feedback used, due to requirements imposed by their learning methods. Such limitations are completely overcome in Netlab.

Giving Adaptive Systems a "Present" €" A new theory of learning is introduced in the book, which embodies and predicts the neurobiological perception of a "present moment". It is characterized by each connection-point having more than one distinct connection-strength. Patented methods included in the described simulation approach fully model the theory.

MultiTemporal Synapses €" A new structure, and accompanying learning method called Weight-To-Weight Learning, is introduced and described. It is based on the above theory of learning. These new mechanisms implement and model the biological experience of a present moment. The advantage is that neural networks are free to quickly and sharply adapt to their present situation, while preserving and enhancing long-term response connections.

Device-Based Neural Network Design and Construction €" Netlab's approach to design is modular, resembling the process by which complex electronic devices are constructed from simpler assemblies. Each assembly, once designed, can be used as a component within the design of other, more complex assemblies.

Introducing Noodle(tm), a Neural Network Description Language €" A chapter introduces, and briefly describes Netlab's new hardware-description programming language. It allows designers to divide-and-conquer complexity, via component libraries and the device-based approach described above.