We are excited to announce the re-launch of our bi-weekly public event, the AI & Pizza Talk Series, where you can get your slice of the latest AI and machine learning research in Cambridge!
Join us at 5:30 pm on 7th June 2023 for an engaging evening featuring two 15-minute talks on cutting-edge research in AI and ML from both academia and industry. After the talks, we will be providing free pizza and refreshments following the talks.
Stay tuned for more information, and we look forward to seeing you at the event!
Location: The small lecture theatre, 21 Station Rd
Time: 17:30 – 18:00 (talks), 18:00 – 19:00 (pizza).
Speakers for the first event:
17:30 – 17:45:
Speaker: Isaac Reid, Machine learning group, University of Cambridge
Title: Simplex Random Features (ICML 23 Oral)
Abstract: Though powerful and mathematically principled, kernel methods notoriously suffer from poor scalability. This has motivated a host of techniques to approximate the kernel’s evaluation, including random features: low-rank decompositions of the Gram matrix constructed using Monte Carlo methods. There is extensive empirical evidence that inducing correlations between the random variates — namely, conditioning that they are orthogonal — can suppress the kernel estimator variance, improving downstream performance in learning tasks. Here, we derive novel analytic closed forms to explain this phenomenon. We also propose a new class, coined simplex random features, which is even better (in fact, optimal among a broad family). We use our new class to approximate approximation in scalable Transformers, obtaining sometimes substantial gains (0.5% on ImageNet) at essentially no cost.
TLDR: first rigorous explanation of the efficacy of orthogonal random features and derivation of a better (optimal) class.
17:45 – 18:00:
Speaker: Jannes Gladrow, Cloud systems future, Microsoft Research Cambridge
Title: Efficient data transport over multimode light-pipes with Megapixel images using differentiable ray tracing and Machine-learning
Abstract: Retrieving images transmitted through multi-mode fibers is of growing interest, thanks to their ability to confine and transport light efficiently in a compact system. Here, we demonstrate Machine-learning-based decoding of large-scale digital images (pages), maximizing page capacity. Using a millimeter-sized square cross-section waveguide, we image a megapixel 8-bit spatial light modulator, presenting data as a matrix of symbols. Normally, (deep-learning) decoders will incur a prohibitive O(n^2) computational scaling to decode n symbols in spatially scrambled data. However, by combining a digital twin of the setup with a U-Net, we can retrieve up to 66 kB using efficient convolutional operations only. We compare trainable ray-tracing-based with eigenmode-based twins and show the former to be superior thanks to its ability to overcome the simulation-to-experiment gap by adjusting to optical imperfections. We train the pipeline end-to-end using a differentiable mutual-information estimator based on the von-Mises distribution which is generally applicable to phase-coding channels.