You can read the full report here

Abstract

On the extreme end of low precision neural networks are binary neural networks (BNN), where weights and activations are quantized to 1 and -1. BNNs have the potential to run deep networks on dramatically reduced hardware. However, BNNs often need special training methods making transfer learning difficult. Furthermore, it is cumbersome to require multiple versions of the same networks for different domains. To alleviate this, adapters are investigated as a solution to this problem.

Adapters inject additional parameters into the network and are a fraction of the size of the base network. During finetuning, the weights in the BNN are kept frozen and the adapters are trained to repurpose the BNN to the new domain. This eliminates the need to have multiple copies of the same deep network. Instead, adapters can be swapped out to repurpose the base network to a new domain. Because of their smaller memory footprint, multiple adapters can be stored for the same cost as a single deep network to cover multiple tasks.

In this project, three points are examined; the effectiveness of adapters compared to standard transfer learning methods, the training time and finally the size of an adapter needed to maintain high accuracy. Both a serial and parallel adapter approach were considered.

Convolutional based BNNs are examined with the task being focused on image classification. Using pruning and grouped convolutions to reduce weights, experiments indicate that adapters that are between 7-12% of the memory size of the base BNN provide a good compromise between accuracy and low memory footprint. Accuracy of BNNs are only a few percentage points off standard transfer learning methods while providing faster training times and flexibility in design approach. This project opens a path to implementing deep networks on low end hardware via BNNs with adapters being used to tailor the network to the specific task.

You can read rest here