Invited paper: Resource sharing in feed forward neural networks for energy efficiency

Exploiting resource reusability and low precision in neural networks is a promising approach to achieve energy efficient computational platforms. This research presents two generalizable approaches to reuse resources in feed-forward neural networks and demonstrated on extreme learning machines. In the first approach, coalescing, a single stack of neuronal units perform both feature extraction and classification tasks through shared resources. In the second approach, folding, the neurons in a high-dimension feedforward layer are folded to execute multiple-tasks. The folding technique can also be combined with low precision modules. The proposed design techniques are validated for a classification task on binary (Australian credit and Diabetes corpus) and multi-class (MNIST) dataset. The total power consumption is measured to be 3.65 mW on TSMC 65nm technology node, while yielding an accuracy of 91.7% for MNIST.