Towards an artificial neural network framework

This paper proposes a framework for hardware artificial neural networks (ANN) combining scalability with the flexibility of software solutions and the speed of hardware ANNs. Our implementation consists of analog neural network blocks realized as ASICs configurable to form arbitrary and large networks having simple elementary resources, i.e. synapses and neurons. Scalability is assured by confining the analog processing of the synapses to blocks and using digital signalling between them. With the help of a genetic algorithm we train the network to combine its elementary resources to form variable network building blocks. We demonstrate how three binary input neurons can act as a single 3-bit neuron and how a group of neurons and synapses can be trained to form a 3-bit output neuron with linear and sigmoid activation functions.

[1]  Alexandre Schmid,et al.  Hardware realization of a Hamming neural network with on-chip learning , 1998, ISCAS '98. Proceedings of the 1998 IEEE International Symposium on Circuits and Systems (Cat. No.98CH36187).

[2]  Johannes Schemmel,et al.  Exploring The Parameter Space Of A Genetic Algorithm For Training An Analog Neural Network , 2002, GECCO.

[3]  Maurizio Valle,et al.  Perspectives on dedicated hardware implementations , 2001, ESANN.

[4]  Johannes Schemmel,et al.  An integrated mixed-mode neural network architecture for megasynapse ANNs , 2002, Proceedings of the 2002 International Joint Conference on Neural Networks. IJCNN'02 (Cat. No.02CH37290).

[5]  Johannes Schemmel,et al.  A VLSI Implementation of an Analog Neural Network Suited for Genetic Algorithms , 2001, ICES.

[6]  Clark S. Lindsey,et al.  Survey of neural network hardware , 1995, SPIE Defense + Commercial Sensing.

[7]  G.M. Bo,et al.  A VLSI architecture for weight perturbation on chip learning implementation , 2000, Proceedings of the IEEE-INNS-ENNS International Joint Conference on Neural Networks. IJCNN 2000. Neural Computing: New Challenges and Perspectives for the New Millennium.