Outlier Features (OFs) are neurons whose activation magnitudes significantly exceed the average over a neural network's (NN) width. They are well known to emerge during standard transformer training and have the undesirable effect of hindering quantisation in afflicted models. Despite their practical importance, little is known behind why OFs emerge during training, nor how one can minimise them.Our work focuses on the above questions, first identifying several quantitative metrics, such as the kurtosis over neuron activation norms, to measure OFs. With these metrics, we study how architectural and optimisation choices influence OFs, and provide practical insights to minimise OFs during training. As highlights, we introduce a novel unnormalised transformer block, the Outlier Protected block, and present a previously unknown benefit of non-diagonal preconditioning optimisers, finding both approaches to significantly reduce OFs and improve quantisation without compromising convergence speed, at scales of up to 7B parameters. Notably, our combination of OP block and non-diagonal preconditioner (SOAP) achieves 14.87 weight-and-activation int8 perplexity (from 14.71 in standard precision), compared to 63.4 int8 perplexity (from 16.00) with a default OF-prone combination of Pre-Norm model and Adam, when quantising OPT-125m models post-training.

Weight-sharing quantization has emerged as a technique to reduce energy expenditure during inference in large neural networks by constraining their weights to a limited set of values. However, existing methods often assume weights are treated solely based on value, neglecting the unique role of weight position. This paper proposes a probabilistic framework based on Bayesian neural networks (BNNs) and a variational relaxation to identify which weights can be moved to which cluster center and to what degree based on their individual position-specific learned uncertainty distributions. We introduce a new initialization setting and a regularization term, enabling the training of BNNs with complex dataset-model combinations. Leveraging the flexibility of weight values from probability distributions, we enhance noise resilience and compressibility. Our iterative clustering procedure demonstrates superior compressibility and higher accuracy compared to state-of-the-art methods on both ResNet models and the more complex transformer-based architectures. In particular, our method outperforms the state-of-the-art quantization method top-1 accuracy by 1.6\% on ImageNet using DeiT-Tiny, with its 5 million+ weights now represented by only 296 unique values.

Approximate Nearest Neighbour (ANN) search is a fundamental problem in information retrieval, underpinning large-scale applications in computer vision, natural language processing, and cross-modal search. Hashing-based methods provide an efficient solution by mapping high-dimensional data into compact binary codes that enable fast similarity computations in Hamming space. Over the past two decades, a substantial body of work has explored learning to hash, where projection and quantisation functions are optimised from data rather than chosen at random. This article offers a foundational survey of early learning-based hashing methods, with an emphasis on the core ideas that shaped the field. We review supervised, unsupervised, and semi-supervised approaches, highlighting how projection functions are designed to generate meaningful embeddings and how quantisation strategies convert these embeddings into binary codes. We also examine extensions to multi-bit and multi-threshold models, as well as early advances in cross-modal retrieval. Rather than providing an exhaustive account of the most recent methods, our goal is to introduce the conceptual foundations of learning-based hashing for ANN search. By situating these early models in their historical context, we aim to equip readers with a structured understanding of the principles, trade-offs, and open challenges that continue to inform current research in this area.

Weight quantisation is an essential technique for enabling efficient training and deployment of modern deep learning models. However, the recipe book of quantisation formats is large and formats are often chosen empirically. In this paper, we propose a framework for systematic design and analysis of quantisation formats. By connecting the question of format design with the classical quantisation theory, we show that the strong practical performance of popular formats comes from their ability to represent values using variable-length codes. We frame the problem as minimising the KL divergence between original and quantised model outputs under a model size constraint, which can be approximated by minimising the squared quantisation error, a well-studied problem where entropy-constrained quantisers with variable-length codes are optimal. We develop non-linear quantisation curves for block-scaled data across multiple distribution families and observe that these formats, along with sparse outlier formats, consistently outperform fixed-length formats, indicating that they also exploit variable-length encoding. Finally, by using the relationship between the Fisher information and KL divergence, we derive the optimal allocation of bit-widths to individual parameter tensors across the model's layers, saving up to 0.25 bits per parameter when applied to large language models. Weight quantisation enables large deep learning models to run on low-resource hardware and edge devices by saving space and memory bandwidth usage. It can be seen as an optimisation problem, where the goal is to retain the behaviour of the high-precision reference model while reducing the total number of bits needed to store its parameters. This naturally splits into two sub-problems of format design and quantisation procedure, both of which are highly active areas of research. We focus on the format design question, i.e., how to choose a representation space for model parameters. This is somewhat independent from the quantisation procedure, which aims to find an optimal point in that space.

Global quantisation ( Q3): To prove ( Q3), we start by giving two auxiliary inequalities.

This setting models data-parallel optimisation, and covers virtually all practical settings, from large-scale regression, to the training of deep neural networks.

Many modern search domains comprise high-dimensional vectors of floating point numbers derived from neural networks, in the form of embeddings. Typical embeddings range in size from hundreds to thousands of dimensions, making the size of the embeddings, and the speed of comparison, a significant issue. Quantisation is a class of mechanism which replaces the floating point values with a smaller representation, for example a short integer. This gives an approximation of the embedding space in return for a smaller data representation and a faster comparison function. Here we take this idea almost to its extreme: we show how vectors of arbitrary-precision floating point values can be replaced by vectors whose elements are drawn from the set {-1,0,1}. This yields very significant savings in space and metric evaluation cost, while maintaining a strong correlation for similarity measurements. This is achieved by way of a class of convex polytopes which exist in the high-dimensional space. In this article we give an outline description of these objects, and show how they can be used for the basis of such radical quantisation while maintaining a surprising degree of accuracy.

Outlier Features (OFs) are neurons whose activation magnitudes significantly exceed the average over a neural network's (NN) width. They are well known to emerge during standard transformer training and have the undesirable effect of hindering quantisation in afflicted models. Despite their practical importance, little is known behind why OFs emerge during training, nor how one can minimise them.Our work focuses on the above questions, first identifying several quantitative metrics, such as the kurtosis over neuron activation norms, to measure OFs. With these metrics, we study how architectural and optimisation choices influence OFs, and provide practical insights to minimise OFs during training. As highlights, we introduce a novel unnormalised transformer block, the Outlier Protected block, and present a previously unknown benefit of non-diagonal preconditioning optimisers, finding both approaches to significantly reduce OFs and improve quantisation without compromising convergence speed, at scales of up to 7B parameters.

While offering a principled framework for uncertainty quantification in deep learning, the employment of Bayesian Neural Networks (BNNs) is still constrained by their increased computational requirements and the convergence difficulties when training very deep, state-of-the-art architectures. In this work, we reinterpret weight-sharing quantization techniques from a stochastic perspective in the context of training and inference with Bayesian Neural Networks (BNNs). Specifically, we leverage 2D adaptive Gaussian distributions, Wasserstein distance estimations, and alpha blending to encode the stochastic behaviour of a BNN in a lower dimensional, soft Gaussian representation. Through extensive empirical investigation, we demonstrate that our approach significantly reduces the computational overhead inherent in Bayesian learning by several orders of magnitude, enabling the efficient Bayesian training of large-scale models, such as ResNet-101 and Vision Transformer (VIT). On various computer vision benchmarks including CIFAR10, CIFAR100, and ImageNet1k. Our approach compresses model parameters by approximately 50x and reduces model size by 75, while achieving accuracy and uncertainty estimations comparable to the state-of-the-art.