Quantization in Deep Learning

Deep learning has a growing history of successes, but heavy algorithms running on large graphical processing units are far from ideal. A relatively new family of deep learning methods called quantized neural networks have appeared in answer to this discrepancy. In Leapmind R&D, we are working on quantization methods, among others, for enabling efficient high-performance deep learning computation on small devices. Neural networks are composed of multiple layers of parameters, each layer transforms the input image, separating and contracting [0] the feature space, resulting in the separation of input images to their various classes. Perhaps the most notable of deep learning problems are image classification, object detection, and segmentation.