Large language model training with 8-bit floating point (FP8) formats promises significant efficiency improvements, but reduced numerical precision makes training challenging. It is currently possible to train in FP8 only if one is willing to tune various hyperparameters, reduce model scale, or accept the overhead of computing dynamic scale factors. We demonstrate simple, scalable FP8 training that requires no dynamic scaling factors or special hyperparameters, even at large model sizes. Our method, \textit{µnit Scaling (µS)}, also enables simple hyperparameter transfer across model widths, matched numerics across training and inference, and other desirable properties. µnit Scaling is straightforward to implement, consisting of a set of minimal interventions based on a first-principles analysis of transformer operations. We validate our method by training models with parameters ranging from 1B to 13B, performing all hidden linear layer computations in FP8. We achieve quality equal to higher-precision baselines while also training up to 33% faster.

Although deep learning has made great progress in recent years, the exploding economic and environmental costs of training neural networks are becoming unsustainable. To address this problem, there has been a great deal of research on *algorithmically-efficient deep learning*, which seeks to reduce training costs not at the hardware or implementation level, but through changes in the semantics of the training program. In this paper, we present a structured and comprehensive overview of the research in this field. First, we formalize the *algorithmic speedup* problem, then we use fundamental building blocks of algorithmically efficient training to develop a taxonomy. Our taxonomy highlights commonalities of seemingly disparate methods and reveals current research gaps. Next, we present evaluation best practices to enable comprehensive, fair, and reliable comparisons of speedup techniques. To further aid research and applications, we discuss common bottlenecks in the training pipeline (illustrated via experiments) and offer taxonomic mitigation strategies for them. Finally, we highlight some unsolved research challenges and present promising future directions.

Shortcut learning, in which models make use of easy-to-represent but unstable associations, is a major failure mode for robust machine learning. We study a flexible, causally-motivated approach to training robust predictors by discouraging the use of specific shortcuts, focusing on a common setting where a robust predictor could achieve optimal i.i.d generalization in principle, but is overshadowed by a shortcut predictor in practice. Our approach uses auxiliary labels, typically available at training time, to enforce conditional independences implied by the causal graph. We show both theoretically and empirically that causally-motivated regularization schemes (a) lead to more robust estimators that generalize well under distribution shift, and (b) have better finite sample efficiency compared to usual regularization schemes, even when no shortcut is present. Our analysis highlights important theoretical properties of training techniques commonly used in the causal inference, fairness, and disentanglement literatures. Our code is available at github.com/mymakar/causally_motivated_shortcut_removal

Test-time augmentation—the aggregation of predictions across transformed versions of a test input—is a common practice in image classification. Traditionally, predictions are combined using a simple average. In this paper, we present 1) experimental analyses that shed light on cases in which the simple average is suboptimal and 2) a method to address these shortcomings. A key finding is that even when test-time augmentation produces a net improvement in accuracy, it can change many correct predictions into incorrect predictions. We delve into when and why test-time augmentation changes a prediction from being correct to incorrect and vice versa. Building on these insights, we present a learning-based method for aggregating test-time augmentations. Experiments across a diverse set of models, datasets, and augmentations show that our method delivers consistent improvements over existing approaches.

Multiplying matrices is among the most fundamental and most computationally demanding operations in machine learning and scientific computing. Consequently, the task of efficiently approximating matrix products has received significant attention. We introduce a learning-based algorithm for this task that greatly outperforms existing methods. Experiments using hundreds of matrices from diverse domains show that it often runs 10x faster than alternatives at a given level of error, as well as 100x faster than exact matrix multiplication. In the common case that one matrix is known ahead of time, our method also has the interesting property that it requires zero multiply-adds. These results suggest that a mixture of hashing, averaging, and byte shuffling{—}the core operations of our method{—}could be a more promising building block for machine learning than the sparsified, factorized, and/or scalar quantized matrix products that have recently been the focus of substantial research and hardware investment.

Vectors of data are at the heart of machine learning and data mining. Recently, vector quantization methods have shown great promise in reducing both the time and space costs of operating on vectors. We introduce a vector quantization algorithm that can compress vectors over 12x faster than existing techniques while also accelerating approximate vector operations such as distance and dot product computations by up to 10x. Because it can encode over 2GB of vectors per second, it makes vector quantization cheap enough to employ in many more circumstances. For example, using our technique to compute approximate dot products in a nested loop can multiply matrices faster than a state-of-the-art BLAS implementation, even when our algorithm must first compress the matrices. In addition to showing the above speedups, we demonstrate that our approach can accelerate nearest neighbor search and maximum inner product search by over 100x compared to floating point operations and 10x compared to other vector quantization methods. Our approximate Euclidean distance and dot product computations are not only faster than those of related algorithms with slower encodings, but also faster than Hamming distance computations, which have direct hardware support on the tested platforms. We also assess the errors of our algorithm's approximate distances and dot products, and find that it is competitive with existing, slower vector quantization algorithms.

Thanks to the rise of wearable and connected devices, sensor-generated time series comprise a large and growing fraction of the world's data. Unfortunately, extracting value from this data can be challenging, since sensors report low-level signals (e.g., acceleration), not the high-level events that are typically of interest (e.g., gestures). We introduce a technique to bridge this gap by automatically extracting examples of real-world events in low-level data, given only a rough estimate of when these events have taken place. By identifying sets of features that repeat in the same temporal arrangement, we isolate examples of such diverse events as human actions, power consumption patterns, and spoken words with up to 96% precision and recall. Our method is fast enough to run in real time and assumes only minimal knowledge of which variables are relevant or the lengths of events. Our evaluation uses numerous publicly available datasets and over 1 million samples of manually labeled sensor data.

Intelligent organisms face a variety of tasks requiring the acquisition of expertise within a specific domain, including the ability to discriminate between a large number of similar patterns. From an energy-efficiency perspective, effective discrimination requires a prudent allocation of neural resources with more frequent patterns and their variants being represented with greater precision. In this work, we demonstrate a biologically plausible means of constructing a single-layer neural network that adaptively (i.e., without supervision) meets this criterion. Specifically, the adaptive algorithm includes synaptogenesis, synaptic shedding, and bi-directional synaptic weight modification to produce a network with outputs (i.e. neural codes) that represent input patterns proportional to the frequency of related patterns. In addition to pattern frequency, the correlational structure of the input environment also affects allocation of neural resources. The combined synaptic modification mechanisms provide an explanation of neuron allocation in the case of self-taught experts.