As Large Language Models (LLMs) grow in size and context length, efficient inference strategies are essential to maintain low-latency token generation. Unfortunately, conventional tensor and data parallelism face diminishing returns when scaling across multiple devices. We propose a novel form—attention-level speculative parallelism (ALSpec)—that predicts self-attention outputs to execute subsequent operations early on separate devices. Our approach overlaps attention and non-attention computations, reducing the attention latency overhead at 128K context length by up to 5x and improving end-to-end decode latency by up to 1.65x, all without sacrificing quality. We establish the fundamental pillars for speculative execution and provide an execution paradigm that simplifies implementation. We show that existing attention-approximation methods perform well on simple information retrieval tasks, but they fail in advanced reasoning and math. Combined with speculative execution, we can approximate up to 90% of self-attention without harming model correctness. Demonstrated on Tenstorrent’s NPU devices, we scale up LLM inference beyond current techniques, paving the way for faster inference in transformer models.

We present a novel approach to the 3D sound source localization task for distributed ad-hoc microphone arrays by formulating it as a set-to-set regression problem. By training a multi-modal masked autoencoder model that operates on audio recordings and microphone coordinates, we show that such a formulation allows for accurate localization of the sound source, by reconstructing coordinates masked in the input. Our approach is flexible in the sense that a single model can be used with an arbitrary number of microphones, even when a subset of audio recordings and microphone coordinates are missing. We test our method on simulated and real-world recordings of music and speech in indoor environments, and demonstrate competitive performance compared to both classical and other learning based localization methods.

While the Transformer architecture has become ubiquitous in the machine learning field, its adaptation to 3D shape recognition is non-trivial. Due to its quadratic computational complexity, the self-attention operator quickly becomes inefficient as the set of input points grows larger. Furthermore, we find that the attention mechanism struggles to find useful connections between individual points on a global scale. In order to alleviate these problems, we propose a two-stage Point Transformer-in-Transformer (Point-TnT) approach which combines local and global attention mechanisms, enabling both individual points and patches of points to attend to each other effectively. Experiments on shape classification show that such an approach provides more useful features for downstream tasks than the baseline Transformer, while also being more computationally efficient. In addition, we also extend our method to feature matching for scene reconstruction, showing that it can be used in conjunction with existing scene reconstruction pipelines.

Speaker localization using microphone arrays depends on accurate time delay estimation techniques. For decades, methods based on the generalized cross correlation with phase transform (GCC-PHAT) have been widely adopted for this purpose. Recently, the GCC-PHAT has also been used to provide input features to neural networks in order to remove the effects of noise and reverberation, but at the cost of losing theoretical guarantees in noise-free conditions. We propose a novel approach to extending the GCC-PHAT, where the received signals are filtered using a shift equivariant neural network that preserves the timing information contained in the signals. By extensive experiments we show that our model consistently reduces the error of the GCC-PHAT in adverse environments, with guarantees of exact time delay recovery in ideal conditions.

The Transformer architecture has been successful across many domains, including natural language processing, computer vision and speech recognition. In keyword spotting, self-attention has primarily been used on top of convolutional or recurrent encoders. We investigate a range of ways to adapt the Transformer architecture to keyword spotting and introduce the Keyword Transformer (KWT), a fully self-attentional architecture that exceeds state-of-the-art performance across multiple tasks without any pre-training or additional data. Surprisingly, this simple architecture outperforms more complex models that mix convolutional, recurrent and attentive layers. KWT can be used as a drop-in replacement for these models, setting two new benchmark records on the Google Speech Commands dataset with 98.6% and 97.7% accuracy on the 12 and 35-command tasks respectively.

Regression via classification (RvC) is a common method used for regression problems in deep learning, where the target variable belongs to a set of continuous values. By discretizing the target into a set of non-overlapping classes, it has been shown that training a classifier can improve neural network accuracy compared to using a standard regression approach. However, it is not clear how the set of discrete classes should be chosen and how it affects the overall solution. In this work, we propose that using several discrete data representations simultaneously can improve neural network learning compared to a single representation. Our approach is end-to-end differentiable and can be added as a simple extension to conventional learning methods, such as deep neural networks. We test our method on three challenging tasks and show that our method reduces the prediction error compared to a baseline RvC approach while maintaining a similar model complexity.

We present PolyLens, a new collaborative filtering recommender system designed to recommend items for groups of users, rather than for individuals. A group recommender is more appropriate and useful for domains in which several people participate in a single activity, as is often the case with movies and restaurants. We present an analysis of the primary design issues for group recommenders, including questions about the nature of groups, the rights of group members, social value functions for groups, and interfaces for displaying group recommendations. We then report on our PolyLens prototype and the lessons we learned from usage logs and surveys from a nine-month trial that included 819 users. We found that users not only valued group recommendations, but were willing to yield some privacy to get the benefits of group recommendations. Users valued an extension to the group recommender system that enabled them to invite non-members to participate, via email.