Streaming Batch Gradient Tracking for Neural Network Training (Student Abstract)

Abstract Faster and more energy efficient hardware accelerators are critical for machine learning on very large datasets. The energy cost of performing vector-matrix multiplication and repeatedly moving neural network models in and out of memory motivates a search for alternative hardware and algorithms. We propose to use streaming batch principal component analysis (SBPCA) to compress batch data during training by using a rank-k approximation of the total batch update. This approach yields comparable training performance to minibatch gradient descent (MBGD) at the same batch size while reducing overall memory and compute requirements.