In Defense of Pure 16-bit Floating-Point Neural Networks

Reducing the number of bits needed to encode the weights and activations of neural networks is highly desirable as it speeds up their training and inference time while reducing memory consumption. For these reasons, research in this area has attracted significant attention toward developing neural networks that leverage lower-precision computing, such as mixed-precision training. Interestingly, none of the existing approaches has investigated pure 16-bit floating-point settings. In this paper, we shed light on the overlooked efficiency of pure 16-bit floating-point neural networks. As such, we provide a comprehensive theoretical analysis to investigate the factors contributing to the differences observed between 16-bit and 32-bit models. We formalize the concepts of floating-point error and tolerance, enabling us to quantitatively explain the conditions under which a 16-bit model can closely approximate the results of its 32-bit counterpart. This theoretical exploration offers perspective that is distinct from the literature which attributes the success of low-precision neural networks to its regularization effect. This in-depth analysis is supported by an extensive series of experiments. Our findings demonstrate that pure 16-bit floating-point neural networks can achieve similar or even better performance than their mixed-precision and 32-bit counterparts. We believe the results presented in this paper will have significant implications for machine learning practitioners, offering an opportunity to reconsider using pure 16-bit networks in various applications.