Faster Multi-GPU Training with PPLL: A Pipeline Parallelism Framework Leveraging Local Learning

Currently, training large-scale deep learning models is typically achieved through parallel training across multiple GPUs. However, due to the inherent communication overhead and synchronization delays in traditional model parallelism methods, seamless parallel training cannot be achieved, which, to some extent, affects overall training efficiency. To address this issue, we present PPLL (Pipeline Parallelism based on Local Learning), a novel framework that leverages local learning algorithms to enable effective parallel training across multiple GPUs. PPLL divides the model into several distinct blocks, each allocated to a separate GPU. By utilizing queues to manage data transfers between GPUs, PPLL ensures seamless cross-GPU communication, allowing multiple blocks to execute forward and backward passes in a pipelined manner. This design minimizes idle times and prevents bottlenecks typically caused by sequential gradient updates, thereby accelerating the overall training process. We validate PPLL through extensive experiments using ResNet and Vision Transformer (ViT) architectures on CIFAR-10, SVHN, and STL-10 datasets. Our results demonstrate that PPLL significantly enhances the training speed of the local learning method while achieving comparable or even superior training speed to traditional pipeline parallelism (PP) without sacrificing model performance. In a 4-GPU training setup, PPLL accelerated local learning training on ViT and ResNet by 162% and 33%, respectively, achieving 1.25x and 0.85x the speed of traditional pipeline parallelism.

Momentum Auxiliary Network for Supervised Local Learning

Deep neural networks conventionally employ end-to-end backpropagation for their training process, which lacks biological credibility and triggers a locking dilemma during network parameter updates, leading to significant GPU memory use. Supervised local learning, which segments the network into multiple local blocks updated by independent auxiliary networks. However, these methods cannot replace end-to-end training due to lower accuracy, as gradients only propagate within their local block, creating a lack of information exchange between blocks. To address this issue and establish information transfer across blocks, we propose a Momentum Auxiliary Network (MAN) that establishes a dynamic interaction mechanism. The MAN leverages an exponential moving average (EMA) of the parameters from adjacent local blocks to enhance information flow. This auxiliary network, updated through EMA, helps bridge the informational gap between blocks. Nevertheless, we observe that directly applying EMA parameters has certain limitations due to feature discrepancies among local blocks. To overcome this, we introduce learnable biases, further boosting performance. We have validated our method on four image classification datasets (CIFAR-10, STL-10, SVHN, ImageNet), attaining superior performance and substantial memory savings. Notably, our method can reduce GPU memory usage by more than 45\% on the ImageNet dataset compared to end-to-end training, while achieving higher performance. The Momentum Auxiliary Network thus offers a new perspective for supervised local learning. Our code is available at: https://github.com/JunhaoSu0/MAN.

MLAAN: Scaling Supervised Local Learning with Multilaminar Leap Augmented Auxiliary Network

End-to-end (E2E) training approaches are commonly plagued by high memory consumption, reduced efficiency in training, challenges in model parallelization, and suboptimal biocompatibility. Local learning is considered a novel interactive training method that holds promise as an alternative to E2E. Nonetheless, conventional local learning methods fall short in achieving high model accuracy due to inadequate local inter-module interactions. In this paper, we introduce a new model known as the Scaling Supervised Local Learning with Multilaminar Leap Augmented Auxiliary Network (MLAAN). MLAAN features an innovative supervised local learning approach coupled with a robust reinforcement module. This dual-component design enables the MLAAN to integrate smoothly with established local learning techniques, thereby enhancing the efficacy of the foundational methods. The method simultaneously acquires the local and global features of the model separately by constructing an independent auxiliary network and a cascade auxiliary network on the one hand and incorporates a leap augmented module, which serves to counteract the reduced learning capacity often associated with weaker supervision. This architecture not only augments the exchange of information amongst the local modules but also effectively mitigates the model's tendency toward myopia. The experimental evaluations conducted on four benchmark datasets, CIFAR-10, STL-10, SVHN, and ImageNet, demonstrate that the integration of MLAAN with existing supervised local learning methods significantly enhances the original methodologies. Of particular note, MLAAN enables local learning methods to comprehensively outperform end-to-end training approaches in terms of optimal performance while saving GPU memory.

GLCAN: Global-Local Collaborative Auxiliary Network for Local Learning

Traditional deep neural networks typically use end-to-end backpropagation, which often places a big burden on GPU memory. Another promising training method is local learning, which involves splitting the network into blocks and training them in parallel with the help of an auxiliary network. Local learning has been widely studied and applied to image classification tasks, and its performance is comparable to that of end-to-end method. However, different image tasks often rely on different feature representations, which is difficult for typical auxiliary networks to adapt to. To solve this problem, we propose the construction method of Global-Local Collaborative Auxiliary Network (GLCAN), which provides a macroscopic design approach for auxiliary networks. This is the first demonstration that local learning methods can be successfully applied to other tasks such as object detection and super-resolution. GLCAN not only saves a lot of GPU memory, but also has comparable performance to an end-to-end approach on data sets for multiple different tasks.

Listen, Disentangle, and Control: Controllable Speech-Driven Talking Head Generation

Most earlier investigations on talking face generation have focused on the synchronization of lip motion and speech content. However, human head pose and facial emotions are equally important characteristics of natural human faces. While audio-driven talking face generation has seen notable advancements, existing methods either overlook facial emotions or are limited to specific individuals and cannot be applied to arbitrary subjects. In this paper, we propose a one-shot Talking Head Generation framework (SPEAK) that distinguishes itself from general Talking Face Generation by enabling emotional and postural control. Specifically, we introduce the Inter-Reconstructed Feature Disentanglement (IRFD) method to decouple human facial features into three latent spaces. We then design a face editing module that modifies speech content and facial latent codes into a single latent space. Subsequently, we present a novel generator that employs modified latent codes derived from the editing module to regulate emotional expression, head poses, and speech content in synthesizing facial animations. Extensive trials demonstrate that our method can generate realistic talking head with coordinated lip motions, authentic facial emotions, and smooth head movements. The demo video is available at the anonymous link: https://anonymous.4open.science/r/SPEAK-F56E