Improvements in training speed are needed to develop the next generation of deep learning models. To perform such a massive amount of computation in a reasonable time, it is parallelized across multiple GPU cores. Perhaps the most popular parallelization method is to use a large batch of data in each iteration of SGD, so the gradient computation can be performed in parallel on multiple workers. We aim to enable massive parallelization without performance degradation, as commonly observed.
We aim to improve the resource efficiency of deep learning (e.g., energy, bandwidth) for training and inference. Our focus is decreasing the numerical precision of the neural network model is a simple and effective way to improve their resource efficiency. Nearly all recent deep learning related hardware relies heavily on lower precision math. The benefits are a reduction in the memory required to store the neural network, a reduction in chip area, and a drastic improvement in energy efficiency.
Significant research efforts are being invested in improving Deep Neural Networks (DNNs) via various modifications. However, such modifications often cause an unexplained degradation in the generalization performance DNNs to unseen data. Recent findings suggest that this degradation is caused by changes to the hidden algorithmic bias of the training algorithm and model. This bias determines which solution is selected from all solutions which fit the data. We aim to understand and control this algorithmic bias.
