MARViN – Multiple Arithmetic Resolutions Vacillating in Neural Networks
share
Quantization is a technique for reducing deep neural networks (DNNs) training and inference times, which is crucial for training in resource constrained environments or time critical inference applications. State-of-the-art (SOTA) quantization approaches focus on post-training quantization, i.e. quantization of pre-trained DNNs for speeding up inference. Very little work on quantized training exists, which neither al-lows dynamic intra-epoch precision switches nor em-ploys an information theory based switching heuristic. Usually, existing approaches require full precision refinement afterwards and enforce a global word length across the whole DNN. This leads to suboptimal quantization mappings and resource usage. Recognizing these limits, we introduce MARViN, a new quantized training strategy using information theory-based intra-epoch precision switching, which decides on a per-layer basis which precision should be used in order to minimize quantization-induced information loss. Note that any quantization must leave enough precision such that future learning steps do not suffer from vanishing gradients. We achieve an average speedup of 1.86 compared to a float32 basis while limiting mean accuracy degradation on AlexNet/ResNet to only -0.075
READ FULL TEXT