A high speed flat CORDIC based neuron with multi-level activation function for robust pattern recognition

Implementing Neural Networks in hardware has been a major problem due to the complexity involved in generating non-linear functions. The high hardware costs incurred in real time applications can be substantially reduced by adopting a suitable reuse methodology of the neurons. In addition, neurons with high speed of operation are necessitated to realise hardware efficient real time pattern recognition for images with higher resolution. In this regard, the response time and area of a neuron becomes critical in realising VLSI efficient neural networks. In this paper, the digital architecture of a multiple valued logic neuron has been proposed to realise a neural network implementation for real-time pattern recognition purposes. The proposed neuron uses a multilevel sigmoidal function as the activation function. Flat CORDIC, a new variation of the CORDIC algorithm, has been employed to generate the complex multi-level activation function in a VLSI efficient manner. The proposed neuron operates with a 200 MHz clock and has significant hardware and latency savings when compared to conventional CORDIC based neurons.

[1]  Stanley L. Hurst,et al.  Multiple-Valued Logic—its Status and its Future , 1984, IEEE Transactions on Computers.

[2]  Anthony N. Michel,et al.  Analysis and synthesis of discrete-time neural networks with multilevel threshold functions , 1991, 1991., IEEE International Sympoisum on Circuits and Systems.

[3]  J J Hopfield,et al.  Neurons with graded response have collective computational properties like those of two-state neurons. , 1984, Proceedings of the National Academy of Sciences of the United States of America.

[4]  S. Walther.  A unified algorithm for elementary functions , 1899 .

[5]  R. Lippmann,et al.  An introduction to computing with neural nets , 1987, IEEE ASSP Magazine.

[6]  Daniel Etiemble,et al.  Comparison of binary and multivalued ICs according to VLSI criteria , 1988, Computer.

[7]  Geoffrey E. Hinton,et al.  Learning representations by back-propagating errors , 1986, Nature.

[8]  Carver A. Mead,et al.  Implementing neural architectures using analog VLSI circuits , 1989 .

[9]  Thambipillai Srikanthan,et al.  Area, time, power optimization for radix-2 redundant CORDIC rotation engines , 1999, Smart Materials, Nano-, and Micro- Smart Systems.

[10]  Robert W. Newcomb,et al.  Circuits for multi-level neuron nonlinearities , 1992, [Proceedings 1992] IJCNN International Joint Conference on Neural Networks.

[11]  Bimal Gisuthan..  A unified architecture for flat CORDIC , 2000 .

[12]  M. Wedlake,et al.  A CORDIC implementation of a digital artificial neuron , 1997, 1997 IEEE Pacific Rim Conference on Communications, Computers and Signal Processing, PACRIM. 10 Years Networking the Pacific Rim, 1987-1997.