Cabrel Teguemne Fokam

Cabrel Teguemne Fokam, Marcel Nieveler, Lukas König et al.

The increasing size of deep learning models has made distributed training across multiple devices essential. Synchronous, centralized methods incur large communication and synchronization overheads. Communication efficient algorithms can reduce these overheads, but often require extra buffers, remain sensitive to stragglers or parameter drift. We present LayUp, an asynchronous decentralized SGD method with layer-wise updates. LayUp asynchronously exchanges incremental layer-wise updates during backpropagation. It uses randomized gossip communication, enabling updates to be applied as soon as they are available without buffering. These design choices reduce parameter drift and improve robustness to stragglers. We establish a theoretical upper bound for the gradient bias introduced by layer-wise updates and prove convergence of LayUp. We empirically validate LayUp on vision and language modeling tasks, showing convergence up to ~32% faster in terms of wall-clock time compared to synchronous data parallel training and up to ~27% faster than comparable communication efficient algorithms while maintaining better task performance. This speed-up is partly due to higher model FLOPs utilization, as we demonstrate. By injecting delays into the communication between workers, we show that LayUp remains robust to stragglers while DDP and other methods degrade in performance. Overall, LayUp provides a novel practical, straggler-robust alternative for distributed training without sacrificing accuracy.

David Kappel, Khaleelulla Khan Nazeer, Cabrel Teguemne Fokam et al.

The ubiquitous backpropagation algorithm requires sequential updates through the network introducing a locking problem. In addition, back-propagation relies on the transpose of forward weight matrices to compute updates, introducing a weight transport problem across the network. Locking and weight transport are problems because they prevent efficient parallelization and horizontal scaling of the training process. We propose a new method to address both these problems and scale up the training of large models. Our method works by dividing a deep neural network into blocks and introduces a feedback network that propagates the information from the targets backwards to provide auxiliary local losses. Forward and backward propagation can operate in parallel and with different sets of weights, addressing the problems of locking and weight transport. Our approach derives from a statistical interpretation of training that treats output activations of network blocks as parameters of probability distributions. The resulting learning framework uses these parameters to evaluate the agreement between forward and backward information. Error backpropagation is then performed locally within each block, leading to "block-local" learning. Several previously proposed alternatives to error backpropagation emerge as special cases of our model. We present results on a variety of tasks and architectures, demonstrating state-of-the-art performance using block-local learning. These results provide a new principled framework for training networks in a distributed setting.