Search

Communication compression is a crucial technique for modern distributed learning systems to alleviate their communication bottlenecks over slower networks. Despite recent intensive studies of gradient compression for data parallel-style training, compressing the activations for models trained with pipeline parallelism is still an open problem. In this paper, we propose AQ-SGD, a novel activation compression algorithm for communication-efficient pipeline parallelism training over slow networks. Different from previous efforts in activation compression, instead of compressing activation values directly, AQ-SGD compresses the changes of the activations. This allows us to show, to the best of our knowledge for the first time, that one can still achieve O ( 1 / √ T ) convergence rate for non-convex objectives under activation compression, without making assumptions on gradient unbiasedness that do not hold for deep learning models with non-linear activation functions. We then show that AQ-SGD can be optimized and implemented efficiently, without additional end-to-end runtime overhead. We evaluated AQ-SGD to fine-tune language models with up to 1.5 billion parameters, compressing activation to 2-4 bits. AQ-SGD provides up to 4.3 × end-to-end speed-up in slower networks, without sacrificing model quality. Moreover, we also show that AQ-SGD can be combined with state-of-the-art gradient compression algorithms to enable end-to-end communication compression: All communications between machines, including model gradients, forward activations, and backward gradients are compressed into lower precision. This provides up to 4.9 × end-to-end speed-up, without sacrificing model quality.

Principal component analysis (PCA) is one of the most powerful tools in machine learning. The simplest method for PCA, the power iteration, requires $\mathcal O(1/\Delta)$ full-data passes to recover the principal component of a matrix with eigen-gap $\Delta$. Lanczos, a significantly more complex method, achieves an accelerated rate of $\mathcal O(1/\sqrt{\Delta})$ passes. Modern applications, however, motivate methods that only ingest a subset of available data, known as the stochastic setting. In the online stochastic setting, simple algorithms like Oja's iteration achieve the optimal sample complexity $\mathcal O(\sigma^2/\Delta^2)$. Unfortunately, they are fully sequential, and also require $\mathcal O(\sigma^2/\Delta^2)$ iterations, far from the $\mathcal O(1/\sqrt{\Delta})$ rate of Lanczos. We propose a simple variant of the power iteration with an added momentum term, that achieves both the optimal sample and iteration complexity. In the full-pass setting, standard analysis shows that momentum achieves the accelerated rate, $\mathcal O(1/\sqrt{\Delta})$. We demonstrate empirically that naively applying momentum to a stochastic method, does not result in acceleration. We perform a novel, tight variance analysis that reveals the "breaking-point variance" beyond which this acceleration does not occur. By combining this insight with modern variance reduction techniques, we construct stochastic PCA algorithms, for the online and offline setting, that achieve an accelerated iteration complexity $\mathcal O(1/\sqrt{\Delta})$. Due to the embarassingly parallel nature of our methods, this acceleration translates directly to wall-clock time if deployed in a parallel environment. Our approach is very general, and applies to many non-convex optimization problems that can now be accelerated using the same technique., Comment: 37 pages, 5 figures

Low-precision computation is often used to lower the time and energy cost of machine learning, and recently hardware accelerators have been developed to support it. Still, it has been used primarily for inference - not training. Previous low-precision training algorithms suffered from a fundamental tradeoff: as the number of bits of precision is lowered, quantization noise is added to the model, which limits statistical accuracy. To address this issue, we describe a simple low-precision stochastic gradient descent variant called HALP. HALP converges at the same theoretical rate as full-precision algorithms despite the noise introduced by using low precision throughout execution. The key idea is to use SVRG to reduce gradient variance, and to combine this with a novel technique called bit centering to reduce quantization error. We show that on the CPU, HALP can run up to $4 \times$ faster than full-precision SVRG and can match its convergence trajectory. We implemented HALP in TensorQuant, and show that it exceeds the validation performance of plain low-precision SGD on two deep learning tasks.