Machine Learning Research at Apple Showcase: Spatial Audio & Activation Steering
COM3 Level 2
SR21, COM3 02-60
Abstract:
In this talk we will showcase two recent lines of Machine Learning research at Apple. First, we will explore the peculiarities of spatial audio, introduce spatial librispeech as dataset for spatial audio learning, discuss the challenges of aligning text and spatial audio through contrastive learning, and cover the downstream applications of these algorithms. The second half of the talk will focus on recent advances in activation steering. That is, the modification of intermediate activations of both Large Language and Diffusion models to induce (or prevent) user-defined concepts. We will describe how to define the concepts and present two algorithms to steer the activations: Self-Cond, which is based on the expertise of individual neurones; and Linear-Act, which reformulates the problem as transporting distributions of activations.
Bio:
Dr Miguel Sarabia earned both a M.Eng. in Information Systems Engineering and a Ph.D. in robotics from Imperial College London in 2010 and 2016 respectively. Since 2016 he works at Apple, where he is currently a senior research scientist. He leads a research collaboration across the company on representation learning for spatial audio.