Joseph Antognini

Haiguang Li, Usama Pervaiz, Joseph Antognini et al.

On-device machine learning (ODML) enables powerful edge applications, but power consumption remains a key challenge for resource-constrained devices. To address this, developers often face a trade-off between model accuracy and power consumption, employing either computationally intensive models on high-power cores or pared-down models on low-power cores. Both approaches typically lead to a compromise in user experience (UX). This work focuses on the use of Gated Compression (GC) layer to enhance ODML model performance while conserving power and maximizing cost-efficiency, especially for always-on use cases. GC layers dynamically regulate data flow by selectively gating activations of neurons within the neural network and effectively filtering out non-essential inputs, which reduces power needs without compromising accuracy, and enables more efficient execution on heterogeneous compute cores. These improvements enhance UX through prolonged battery life, improved device responsiveness, and greater user comfort. In this work, we have integrated GC layers into vision and speech domain models including the transformer-based ViT model. Our experiments demonstrate theoretical power efficiency gains ranging from 158x to 30,000x for always-on scenarios. This substantial improvement empowers ODML applications with enhanced UX benefits.

Haiguang Li, Usama Pervaiz, Michał Matuszak et al.

On-device machine learning (ODML) enables intelligent applications on resource-constrained devices. However, power consumption poses a major challenge, forcing a trade-off between model accuracy and power efficiency that often limits model complexity. The previously established Gated Compression (GC) layers offer a solution, enabling power efficiency without sacrificing model performance by selectively gating samples that lack signals of interest. However, their reliance on ground truth labels limits GC layers to supervised tasks. This work introduces the Dynamic Switch Layer (DSL), extending the benefits of GC layers to unsupervised learning scenarios, and maintaining power efficiency without the need for labeled data. The DSL builds upon the GC architecture, leveraging a dynamic pathway selection, and adapting model complexity in response to the innate structure of the data. We integrate the DSL into the SoundStream architecture and demonstrate that by routing up to 80% of samples through a lightweight pass we achieve a 12.3x reduction in the amount of computation performed and a 20.9x reduction in model size. This reduces the on-device inference latency by up to 26.5% and improves power efficiency by up to 21.4% without impacting model performance.

Christopher J. Shallue, Jaehoon Lee, Joseph Antognini et al.

Recent hardware developments have dramatically increased the scale of data parallelism available for neural network training. Among the simplest ways to harness next-generation hardware is to increase the batch size in standard mini-batch neural network training algorithms. In this work, we aim to experimentally characterize the effects of increasing the batch size on training time, as measured by the number of steps necessary to reach a goal out-of-sample error. We study how this relationship varies with the training algorithm, model, and data set, and find extremely large variation between workloads. Along the way, we show that disagreements in the literature on how batch size affects model quality can largely be explained by differences in metaparameter tuning and compute budgets at different batch sizes. We find no evidence that larger batch sizes degrade out-of-sample performance. Finally, we discuss the implications of our results on efforts to train neural networks much faster in the future. Our experimental data is publicly available as a database of 71,638,836 loss measurements taken over the course of training for 168,160 individual models across 35 workloads.

Joseph Antognini, Matt Hoffman, Ron J. Weiss

Texture synthesis techniques based on matching the Gram matrix of feature activations in neural networks have achieved spectacular success in the image domain. In this paper we extend these techniques to the audio domain. We demonstrate that synthesizing diverse audio textures is challenging, and argue that this is because audio data is relatively low-dimensional. We therefore introduce two new terms to the original Grammian loss: an autocorrelation term that preserves rhythm, and a diversity term that encourages the optimization procedure to synthesize unique textures. We quantitatively study the impact of our design choices on the quality of the synthesized audio by introducing an audio analogue to the Inception loss which we term the VGGish loss. We show that there is a trade-off between the diversity and quality of the synthesized audio using this technique. We additionally perform a number of experiments to qualitatively study how these design choices impact the quality of the synthesized audio. Finally we describe the implications of these results for the problem of audio style transfer.