Graph structure learning (GSL), which aims to learn the adjacency matrix for graph neural networks (GNNs), has shown great potential in boosting the performance of GNNs. Most existing GSL works apply a joint learning framework where the estimated adjacency matrix and GNN parameters are optimized for downstream tasks. However, as GSL is essentially a link prediction task, whose goal may largely differ from the goal of the downstream task. The inconsistency of these two goals limits the GSL methods to learn the potential optimal graph structure. Moreover, the joint learning framework suffers from scalability issues in terms of time and space during the process of estimation and optimization of the adjacency matrix. To mitigate these issues, we propose a graph structure refinement (GSR) framework with a pretrain-finetune pipeline. Specifically, The pre-training phase aims to comprehensively estimate the underlying graph structure by a multi-view contrastive learning framework with both intra- and inter-view link prediction tasks. Then, the graph structure is refined by adding and removing edges according to the edge probabilities estimated by the pre-trained model. Finally, the fine-tuning GNN is initialized by the pre-trained model and optimized toward downstream tasks. With the refined graph structure remaining static in the fine-tuning space, GSR avoids estimating and optimizing graph structure in the fine-tuning phase which enjoys great scalability and efficiency. Moreover, the fine-tuning GNN is boosted by both migrating knowledge and refining graphs. Extensive experiments are conducted to evaluate the effectiveness (best performance on six benchmark datasets), efficiency, and scalability (13.8x faster using 32.8% GPU memory compared to the best GSL baseline on Cora) of the proposed model.
Recent years have witnessed the booming of various differentiable optimization algorithms. These algorithms exhibit different execution patterns, and their execution needs massive computational resources that go beyond a single CPU and GPU. Existing differentiable optimization libraries, however, cannot support efficient algorithm development and multi-CPU/GPU execution, making the development of differentiable optimization algorithms often cumbersome and expensive. This paper introduces TorchOpt, a PyTorch-based efficient library for differentiable optimization. TorchOpt provides a unified and expressive differentiable optimization programming abstraction. This abstraction allows users to efficiently declare and analyze various differentiable optimization programs with explicit gradients, implicit gradients, and zero-order gradients. TorchOpt further provides a high-performance distributed execution runtime. This runtime can fully parallelize computation-intensive differentiation operations (e.g. tensor tree flattening) on CPUs / GPUs and automatically distribute computation to distributed devices. Experimental results show that TorchOpt achieves $5.2\times$ training time speedup on an 8-GPU server. TorchOpt is available at: https://github.com/metaopt/torchopt/.
Advances in reinforcement learning (RL) often rely on massive compute resources and remain notoriously sample inefficient. In contrast, the human brain is able to efficiently learn effective control strategies using limited resources. This raises the question whether insights from neuroscience can be used to improve current RL methods. Predictive processing is a popular theoretical framework which maintains that the human brain is actively seeking to minimize surprise. We show that recurrent neural networks which predict their own sensory states can be leveraged to minimise surprise, yielding substantial gains in cumulative reward. Specifically, we present the Predictive Processing Proximal Policy Optimization (P4O) agent; an actor-critic reinforcement learning agent that applies predictive processing to a recurrent variant of the PPO algorithm by integrating a world model in its hidden state. Even without hyperparameter tuning, P4O significantly outperforms a baseline recurrent variant of the PPO algorithm on multiple Atari games using a single GPU. It also outperforms other state-of-the-art agents given the same wall-clock time and exceeds human gamer performance on multiple games including Seaquest, which is a particularly challenging environment in the Atari domain. Altogether, our work underscores how insights from the field of neuroscience may support the development of more capable and efficient artificial agents.
We introduce Breadth-First Pipeline Parallelism, a novel training schedule which optimizes the combination of pipeline and data parallelism. Breadth-First Pipeline Parallelism lowers training time, cost and memory usage by combining a high GPU utilization with a small batch size per GPU, and by making use of fully sharded data parallelism. Experimentally, we observed an increase of up to 43% in training throughput for a 52 billion-parameter model using a small batch size per GPU compared to Megatron-LM, which would reduce the training time and cost by the same amount on a large GPU cluster.
Transformers have become central to recent advances in computer vision. However, training a vision Transformer (ViT) model from scratch can be resource intensive and time consuming. In this paper, we aim to explore approaches to reduce the training costs of ViT models. We introduce some algorithmic improvements to enable training a ViT model from scratch with limited hardware (1 GPU) and time (24 hours) resources. First, we propose an efficient approach to add locality to the ViT architecture. Second, we develop a new image size curriculum learning strategy, which allows to reduce the number of patches extracted from each image at the beginning of the training. Finally, we propose a new variant of the popular ImageNet1k benchmark by adding hardware and time constraints. We evaluate our contributions on this benchmark, and show they can significantly improve performances given the proposed training budget. We will share the code in https://github.com/BorealisAI/efficient-vit-training.
Using 3D CNNs on high resolution medical volumes is very computationally demanding, especially for large datasets like the UK Biobank which aims to scan 100,000 subjects. Here we demonstrate that using 2D CNNs on a few 2D projections (representing mean and standard deviation across axial, sagittal and coronal slices) of the 3D volumes leads to reasonable test accuracy when predicting the age from brain volumes. Using our approach, one training epoch with 20,324 subjects takes 20 - 50 seconds using a single GPU, which two orders of magnitude faster compared to a small 3D CNN. These results are important for researchers who do not have access to expensive GPU hardware for 3D CNNs.
Resistive Random-Access Memory (RRAM) is well-suited to accelerate neural network (NN) workloads as RRAM-based Processing-in-Memory (PIM) architectures natively support highly-parallel multiply-accumulate (MAC) operations that form the backbone of most NN workloads. Unfortunately, NN workloads such as transformers require support for non-MAC operations (e.g., softmax) that RRAM cannot provide natively. Consequently, state-of-the-art works either integrate additional digital logic circuits to support the non-MAC operations or offload the non-MAC operations to CPU/GPU, resulting in significant performance and energy efficiency overheads due to data movement. In this work, we propose NEON, a novel compiler optimization to enable the end-to-end execution of the NN workload in RRAM. The key idea of NEON is to transform each non-MAC operation into a lightweight yet highly-accurate neural network. Utilizing neural networks to approximate the non-MAC operations provides two advantages: 1) We can exploit the key strength of RRAM, i.e., highly-parallel MAC operation, to flexibly and efficiently execute non-MAC operations in memory. 2) We can simplify RRAM's microarchitecture by eliminating the additional digital logic circuits while reducing the data movement overheads. Acceleration of the non-MAC operations in memory enables NEON to achieve a 2.28x speedup compared to an idealized digital logic-based RRAM. We analyze the trade-offs associated with the transformation and demonstrate feasible use cases for NEON across different substrates.
When building state-of-the-art speech translation models, the need for large computational resources is a significant obstacle due to the large training data size and complex models. The availability of pre-trained models is a promising opportunity to build strong speech translation systems efficiently. In a first step, we investigate efficient strategies to build cascaded and end-to-end speech translation systems based on pre-trained models. Using this strategy, we can train and apply the models on a single GPU. While the end-to-end models show superior translation performance to cascaded ones, the application of this technology has a limitation on the need for additional end-to-end training data. In a second step, we proposed an additional similarity loss to encourage the model to generate similar hidden representations for speech and transcript. Using this technique, we can increase the data efficiency and improve the translation quality by 6 BLEU points in scenarios with limited end-to-end training data.
A growing number of Machine Learning Frameworks recently made Deep Learning accessible to a wider audience of engineers, scientists, and practitioners, by allowing straightforward use of complex neural network architectures and algorithms. However, since deep learning is rapidly evolving, not only through theoretical advancements but also with respect to hardware and software engineering, ML frameworks often lose backward compatibility and introduce technical debt that can lead to bottlenecks and sub-optimal resource utilization. Moreover, the focus is in most cases not on deep learning engineering, but rather on new models and theoretical advancements. In this work, however, we focus on engineering, more specifically on the data loading pipeline in the PyTorch Framework. We designed a series of benchmarks that outline performance issues of certain steps in the data loading process. Our findings show that for classification tasks that involve loading many files, like images, the training wall-time can be significantly improved. With our new, modified ConcurrentDataloader we can reach improvements in GPU utilization and significantly reduce batch loading time, up to 12X. This allows for the use of the cloud-based, S3-like object storage for datasets, and have comparable training time as if datasets are stored on local drives.
The demonstrated success of transfer learning has popularized approaches that involve pretraining models from massive data sources and subsequent finetuning towards a specific task. While such approaches have become the norm in fields such as natural language processing, implementation and evaluation of transfer learning approaches for chemistry are in the early stages. In this work, we demonstrate finetuning for downstream tasks on a graph neural network (GNN) trained over a molecular database containing 2.7 million water clusters. The use of Graphcore IPUs as an AI accelerator for training molecular GNNs reduces training time from a reported 2.7 days on 0.5M clusters to 1.2 hours on 2.7M clusters. Finetuning the pretrained model for downstream tasks of molecular dynamics and transfer to a different potential energy surface took only 8.3 hours and 28 minutes, respectively, on a single GPU.
Global localisation from visual data is a challenging problem applicable to many robotics domains. Prior works have shown that neural networks can be trained to map images of an environment to absolute camera pose within that environment, learning an implicit neural mapping in the process. In this work we evaluate the applicability of such an approach to real-world robotics scenarios, demonstrating that by constraining the problem to 2-dimensions and significantly increasing the quantity of training data, a compact model capable of real-time inference on embedded platforms can be used to achieve localisation accuracy of several centimetres. We deploy our trained model onboard a UGV platform, demonstrating its effectiveness in a waypoint navigation task, wherein it is able to localise with a mean accuracy of 9cm at a rate of 6fps running on the UGV onboard CPU, 35fps on an embedded GPU, or 220fps on a desktop GPU. Along with this work we will release a novel localisation dataset comprising simulated and real environments, each with training samples numbering in the tens of thousands.
Pre-trained language models (PLMs) have outperformed other NLP models on a wide range of tasks. Opting for a more thorough understanding of their capabilities and inner workings, researchers have established the extend to which they capture lower-level knowledge like grammaticality, and mid-level semantic knowledge like factual understanding. However, there is still little understanding of their knowledge of higher-level aspects of language. In particular, despite the importance of sociodemographic aspects in shaping our language, the questions of whether, where, and how PLMs encode these aspects, e.g., gender or age, is still unexplored. We address this research gap by probing the sociodemographic knowledge of different single-GPU PLMs on multiple English data sets via traditional classifier probing and information-theoretic minimum description length probing. Our results show that PLMs do encode these sociodemographics, and that this knowledge is sometimes spread across the layers of some of the tested PLMs. We further conduct a multilingual analysis and investigate the effect of supplementary training to further explore to what extent, where, and with what amount of pre-training data the knowledge is encoded. Our overall results indicate that sociodemographic knowledge is still a major challenge for NLP. PLMs require large amounts of pre-training data to acquire the knowledge and models that excel in general language understanding do not seem to own more knowledge about these aspects.
The role of mobile cameras increased dramatically over the past few years, leading to more and more research in automatic image quality enhancement and RAW photo processing. In this Mobile AI challenge, the target was to develop an efficient end-to-end AI-based image signal processing (ISP) pipeline replacing the standard mobile ISPs that can run on modern smartphone GPUs using TensorFlow Lite. The participants were provided with a large-scale Fujifilm UltraISP dataset consisting of thousands of paired photos captured with a normal mobile camera sensor and a professional 102MP medium-format FujiFilm GFX100 camera. The runtime of the resulting models was evaluated on the Snapdragon's 8 Gen 1 GPU that provides excellent acceleration results for the majority of common deep learning ops. The proposed solutions are compatible with all recent mobile GPUs, being able to process Full HD photos in less than 20-50 milliseconds while achieving high fidelity results. A detailed description of all models developed in this challenge is provided in this paper.
Protein structure prediction is a critical problem linked to drug design, mutation detection, and protein synthesis, among other applications. To this end, evolutionary data has been used to build contact maps which are traditionally minimized as energy functions via gradient descent based schemes like the L-BFGS algorithm. In this paper we present what we call the Alternating Metropolis-Hastings (AMH) algorithm, which (a) significantly improves the performance of traditional MCMC methods, (b) is inherently parallelizable allowing significant hardware acceleration using GPU, and (c) can be integrated with the L-BFGS algorithm to improve its performance. The algorithm shows an improvement in energy of found structures of 8.17% to 61.04% (average 38.9%) over traditional MH and 0.53% to 17.75% (average 8.9%) over traditional MH with intermittent noisy restarts, tested across 9 proteins from recent CASP competitions. We go on to map the Alternating MH algorithm to a GPGPU which improves sampling rate by 277x and improves simulation time to a low energy protein prediction by 7.5x to 26.5x over CPU. We show that our approach can be incorporated into state-of-the-art protein prediction pipelines by applying it to both trRosetta2's energy function and the distogram component of Alphafold1's energy function. Finally, we note that specially designed probabilistic computers (or p-computers) can provide even better performance than GPUs for MCMC algorithms like the one discussed here.
We introduce pyGSL, a Python library that provides efficient implementations of state-of-the-art graph structure learning models along with diverse datasets to evaluate them on. The implementations are written in GPU-friendly ways, allowing one to scale to much larger network tasks. A common interface is introduced for algorithm unrolling methods, unifying implementations of recent state-of-the-art techniques and allowing new methods to be quickly developed by avoiding the need to rebuild the underlying unrolling infrastructure. Implementations of differentiable graph structure learning models are written in PyTorch, allowing us to leverage the rich software ecosystem that exists e.g., around logging, hyperparameter search, and GPU-communication. This also makes it easy to incorporate these models as components in larger gradient based learning systems where differentiable estimates of graph structure may be useful, e.g. in latent graph learning. Diverse datasets and performance metrics allow consistent comparisons across models in this fast growing field. The full code repository can be found on https://github.com/maxwass/pyGSL.
Tensor program tuning is a non-convex objective optimization problem, to which search-based approaches have proven to be effective. At the core of the search-based approaches lies the design of the cost model. Though deep learning-based cost models perform significantly better than other methods, they still fall short and suffer from the following problems. First, their feature extraction heavily relies on expert-level domain knowledge in hardware architectures. Even so, the extracted features are often unsatisfactory and require separate considerations for CPUs and GPUs. Second, a cost model trained on one hardware platform usually performs poorly on another, a problem we call cross-hardware unavailability. In order to address these problems, we propose TLP and MTLTLP. TLP is a deep learning-based cost model that facilitates tensor program tuning. Instead of extracting features from the tensor program itself, TLP extracts features from the schedule primitives. We treat schedule primitives as tensor languages. TLP is thus a Tensor Language Processing task. In this way, the task of predicting the tensor program latency through the cost model is transformed into a natural language processing (NLP) regression task. MTL-TLP combines Multi-Task Learning and TLP to cope with the cross-hardware unavailability problem. We incorporate these techniques into the Ansor framework and conduct detailed experiments. Results show that TLP can speed up the average search time by 9.1X and 3.0X on CPU and GPU workloads, respectively, compared to the state-of-the-art implementation. MTL-TLP can achieve a speed-up of 4.7X and 2.9X on CPU and GPU workloads, respectively, using only 7% of the target hardware data.
Cryo Focused Ion-Beam Scanning Electron Microscopy (cryo FIB-SEM) enables three-dimensional and nanoscale imaging of biological specimens via a slice and view mechanism. The FIB-SEM experiments are, however, limited by a slow (typically, several hours) acquisition process and the high electron doses imposed on the beam sensitive specimen can cause damage. In this work, we present a compressive sensing variant of cryo FIB-SEM capable of reducing the operational electron dose and increasing speed. We propose two Targeted Sampling (TS) strategies that leverage the reconstructed image of the previous sample layer as a prior for designing the next subsampling mask. Our image recovery is based on a blind Bayesian dictionary learning approach, i.e., Beta Process Factor Analysis (BPFA). This method is experimentally viable due to our ultra-fast GPU-based implementation of BPFA. Simulations on artificial compressive FIB-SEM measurements validate the success of proposed methods: the operational electron dose can be reduced by up to 20 times. These methods have large implications for the cryo FIB-SEM community, in which the imaging of beam sensitive biological materials without beam damage is crucial.
Recurrent Neural Networks (RNNs) have become the standard modeling technique for sequence data, and are used in a number of novel text-to-speech models. However, training a TTS model including RNN components has certain requirements for GPU performance and takes a long time. In contrast, studies have shown that CNN-based sequence synthesis technology can greatly reduce training time in text-to-speech models while ensuring a certain performance due to its high parallelism. We propose a new text-to-speech system based on deep convolutional neural networks that does not employ any RNN components (recurrent units). At the same time, we improve the generality and robustness of our model through a series of data augmentation methods such as Time Warping, Frequency Mask, and Time Mask. The final experimental results show that the TTS model using only the CNN component can reduce the training time compared to the classic TTS models such as Tacotron while ensuring the quality of the synthesized speech.
Oblique plane microscopy, OPM, is a form of lightsheet microscopy that permits volumetric imaging of biological samples at high temporal and spatial resolution. However, the imaging geometry of OPM, and related variants of light sheet microscopy, distorts the coordinate frame of the presented image sections with respect to real space coordinate frame in which the sample is moved to navigate to regions of interest. This makes live viewing and practical operation of such microscopes difficult. We present an open-source software package that utilises GPU acceleration and multiprocessing to transform the display of OPM imaging data in real time to produce live views that mimic that produced by standard widefield microscopes. Image stacks can be acquired, processed and plotted at rates of several Hz, making live operation of OPMs, and similar microscopes, more user friendly and intuitive.
During image editing, existing deep generative models tend to re-synthesize the entire output from scratch, including the unedited regions. This leads to a significant waste of computation, especially for minor editing operations. In this work, we present Spatially Sparse Inference (SSI), a general-purpose technique that selectively performs computation for edited regions and accelerates various generative models, including both conditional GANs and diffusion models. Our key observation is that users prone to gradually edit the input image. This motivates us to cache and reuse the feature maps of the original image. Given an edited image, we sparsely apply the convolutional filters to the edited regions while reusing the cached features for the unedited areas. Based on our algorithm, we further propose Sparse Incremental Generative Engine (SIGE) to convert the computation reduction to latency reduction on off-the-shelf hardware. With about $1\%$-area edits, SIGE accelerates DDPM by $3.0\times$ on NVIDIA RTX 3090 and $4.6\times$ on Apple M1 Pro GPU, Stable Diffusion by $7.2\times$ on 3090, and GauGAN by $5.6\times$ on 3090 and $5.2\times$ on M1 Pro GPU. Compared to our conference version, we extend SIGE to accommodate attention layers and apply it to Stable Diffusion. Additionally, we offer support for Apple M1 Pro GPU and include more results with large and sequential edits.