BinaryFlex: On-the-Fly Kernel Generation in Binary Convolutional Networks

Vincent W.-S. Tseng, Sourav Bhattachary, Javier Fernández Marqués, Milad Alizadeh, Catherine Tong, Nicholas Donald Lane

Abstract: In this work we present BinaryFlex, a neural network architecture that learns weighting coefficients of predefined orthogonal binary basis instead of the conventional approach of learning directly the convolutional filters. We have demonstrated the feasibility of our approach for complex computer vision datasets such as ImageNet. Our architecture trained on ImageNet is able to achieve top-5 accuracy of 65.7% while being around 2x smaller than binary networks capable of achieving similar accuracy levels. By using deterministic basis, that can be generated on-the-fly very efficiently, our architecture offers a great deal of flexibility in memory footprint when deploying in constrained microcontroller devices.