Optimized Extreme Learning Machine for Big Data Applications Using Python
Radu Dogaru, Ioana Dogaru · 2018 International Conference on Communications (COMM) · 2018
This paper reports an optimized implementation for the extreme learning machine (ELM). Among several modeling languages, Python with optimized linear algebra library support was found to be the best choice. A novel nonlinear function, the absolute value, is proposed and shown to provide the best performance and speed among other traditional nonlinearities. A fixed point implementation with finite resolution is proposed and evaluated, concluding that best accuracy is maintained for 2 bits quantization in the input layer and 8 bits in the output layer. The optimized ELM with hardware-convenient nonlinearities and finite precision is widely suitable for solving big data problems on various computing platforms such as FPGA, low cost microcontrollers, etc. In terms of speed, our optimized ELM solution outperforms state of the art Python implementations, providing at least 3 times faster training and retrieval on similar computing platforms.