In the new paper Training Compute-Optimal Large Language Models, a DeepMind research team posits that current large language models are significantly undertrained and, based on empirical outcomes of over 400 training runs, proposes three predictive approaches for optimally setting model size and training duration.
In the new paper DataMUX: Data Multiplexing for Neural Networks, a Princeton University research team proposes Data Multiplexing (DataMUX). The novel technique enables neural networks to process multiple inputs simultaneously and generate accurate predictions, increasing model throughput with minimal additional memory requirements.
In the new paper Visual Attention Network, a research team from Tsinghua University and Nankai University introduces a novel large kernel attention (LKA) mechanism for an extremely simple and efficient Visual Attention Network (VAN) that significantly outperforms state-of-the-art vision transformers and convolutional neural networks on various computer vision tasks.
A research team from Microsoft and NVIDIA leverages the NVIDIA Megatron-LM and Microsoft’s DeepSpeed to create an efficient and scalable 3D parallel system that combines data, pipeline, and tensor-slicing based parallelism, achieving superior zero-, one-, and few-shot learning accuracies and new state-of-the-art results on NLP benchmarks.
In the new paper A Neural Network Solves and Generates Mathematics Problems by Program Synthesis: Calculus, Differential Equations, Linear Algebra, and More, a research team from MIT, Columbia University, Harvard University and University of Waterloo proposes a neural network that can solve university-level mathematics problems via program synthesis.
In the new paper Sparse is Enough in Scaling Transformers, a research team from the University of Warsaw, Google Research and OpenAI proposes Scaling Transformers, a family of novel transformers that leverage sparse layers to scale efficiently and perform unbatched decoding much faster than original transformers, enabling fast inference on long sequences even with limited memory.
In the new paper Gradients Are Not All You Need, a Google Brain and Radboud University research team discusses a “particularly sinister” chaos-based failure mode that appears in a variety of differentiable circumstances, ranging from recurrent neural networks and numerical physics simulation to training learned optimizers.
A DeepMind research team presents the One Pass ImageNet (OPIN) problem, designed to study the space and compute efficiency of deep learning in a streaming setting with constrained data storage and to develop model training systems where each example is passed to the system only once.
A Microsoft Research India team presents Varuna, a system for training massive deep learning models on commodity networking that eliminates the need for specialized hyperclusters and alleviates the cost, scale, and resource utilization challenges of deep learning model training.
In the new paper Understanding How Encoder-Decoder Architectures Attend, researchers from the University of Washington, Google Blueshift Team and Google Brain Team propose a method for decomposing hidden states over a sequence into temporal- and input-driven components, revealing how attention matrices are formed in encoder-decoder networks.
In the new paper Non-deep Networks, a research team from Princeton University and Intel Labs argues it is possible to achieve high performance with “non-deep” neural networks, presenting ParNet (Parallel Networks), a novel 12-layer architecture that achieves performance competitive with its state-of-the-art deep counterparts.
A Google Research team conducts a systematic exploration comprising more than 4800 experiments on Vision Transformers, MLP-Mixers and ResNets with parameters ranging from 10 million to 10 billion, evaluated on more than 20 downstream image recognition tasks, aiming to capture the nonlinear relationships between performance on upstream and downstream tasks.
A research team proposes ConvMixer, an extremely simple model designed to support the argument that the impressive performance of vision transformers (ViTs) is mainly attributable to their use of patches as the input representation. The study shows that ConvMixer can outperform ViTs, MLP-Mixers and classical vision models.
An Apple research team explores multiple architectures and training procedures to develop a novel multi-speaker and multi-lingual neural TTS system. The study combines speech from 30 speakers from 15 locales in 8 languages, and demonstrates that for the vast majority of voices, such multi-lingual and multi-speaker models can yield better quality than single speaker models.
In a 200+ page paper, Percy Liang, Fei-Fei Li, and over 100 other researchers from the Stanford University Center for Research on Foundation Models (CRFM) systematically describe the opportunities and risks of large-scale pretrained “foundation” models. The unique study aims to provide a clearer understanding of how these models work, when and how they fail, and the various capabilities provided by their emergent properties.
A research team from Università di Firenze, Università di Siena, University of Cambridge and Universitè Côte d’Azur proposes a general approach to explainable artificial intelligence (XAI) in neural architectures, designing interpretable deep learning models called Logic Explained Networks (LENs). The novel approach yields better performance than established white-box models while providing more compact and meaningful explanations.
A research team from the University of Science and Technology of China, Microsoft Cloud AI, City University of Hong Kong and Wormpex AI Research propose a robust and invisible backdoor attack called “Poison Ink” and demonstrates its immunity to state-of-the-art defence techniques.
A research team from Zhejiang University, Wuhan University and Adobe Research proposes Feature Importance-Aware Attacks (FIA) that drastically improve the transferability of adversarial examples, achieving superior performance compared to state-of-the-art transferable attacks.
A DeepMind research team proposes Perceiver IO, a single network that can easily integrate and transform arbitrary information for arbitrary tasks while scaling linearly with both input and output sizes. The general architecture achieves outstanding results on tasks with highly structured output spaces, such as natural language and visual understanding.
A research team from Google Research and Northwestern University presents polynomial time and sample-efficient algorithms for learning an unknown depth-2 feedforward neural network with general ReLU activations, aiming to provide insights into whether efficient algorithms exist for learning ReLU networks.
A research team from Facebook AI and UC Berkeley finds a solution for vision transformers’ optimization instability problem by simply using a standard, lightweight convolutional stem for ViT models. The approach dramatically increases optimizer stability and improves peak performance without sacrificing computation efficiency.
Researchers from Google conduct a survey on how to make Deep Learning models smaller, faster, and better. The team focuses on core areas of model efficiency, from modelling techniques to hardware support, and open-sources an experiment-based guide and code to help practitioners optimize their model training and deployment.
A research team from Mila, McGill University, Université de Montréal, DeepMind and Microsoft proposes GFlowNet, a novel flow network-based generative method that can turn a given positive reward into a generative policy that samples with a probability proportional to the return.v
An IEEE team provides a comprehensive overview of the bottom-up and top-down design approaches toward neuromorphic intelligence, highlighting the different levels of granularity present in existing silicon implementations and assessing the benefits of the different circuit design styles in neural processing systems.
A research team from Google proposes GSPMD, an automatic parallelism system for ML computation graphs that uses simple tensor sharding annotations to achieve different parallelism paradigms in a unified way, including data parallelism, within-layer model parallelism, spatial partitioning, weight-update sharding, optimizer-state sharding and pipeline parallelism.
A research team from ETH Zurich leverages existing spike-based learning circuits to propose a biologically plausible architecture that is highly successful in classifying distinct and complex spatio-temporal spike patterns. The work contributes to the design of ultra-low-power mixed-signal neuromorphic processing systems capable of distinguishing spatio-temporal patterns in spiking activity.