BEGIN:VCALENDAR VERSION:2.0 PRODID:-//Talks.cam//talks.cam.ac.uk// X-WR-CALNAME:Talks.cam BEGIN:VEVENT SUMMARY:Microsoft AI &\; Pizza event - Speaker to be confirmed DTSTART:20230607T163000Z DTEND:20230607T180000Z UID:TALK202165@talks.cam.ac.uk CONTACT:Kimberly Cole DESCRIPTION:We are excited to announce the re-launch of our bi-weekly publ ic event\, the AI & Pizza Talk Series\, where you can get your slice of th e latest AI and machine learning research in Cambridge!\n \nJoin us at 5:3 0 pm on 7th June 2023 for an engaging evening featuring two 15-minute talk s on cutting-edge research in AI and ML from both academia and industry. A fter the talks\, we will be providing free pizza and refreshments followin g the talks. \n \nStay tuned for more information\, and we look forward to seeing you at the event!\n \nCheers\,\nChao\n \nLocation: The small lectu re theatre\, 21 Station Rd\n \nTime: 17:30 – 18:00 (talks)\, 18:00 – 1 9:00 (pizza).\n \n \nSpeakers for the first event:\n \n17:30 – 17:45:\n \nSpeaker: Isaac Reid\, Machine learning group\, University of Cambridge\n \nTitle: Simplex Random Features (ICML 23 Oral)\n \nAbstract: Though powe rful and mathematically principled\, kernel methods notoriously suffer fro m poor scalability. This has motivated a host of techniques to approximate the kernel's evaluation\, including random features: low-rank decompositi ons of the Gram matrix constructed using Monte Carlo methods. There is ext ensive empirical evidence that inducing correlations between the random va riates -- namely\, conditioning that they are orthogonal -- can suppress t he kernel estimator variance\, improving downstream performance in learnin g tasks. Here\, we derive novel analytic closed forms to explain this phen omenon. We also propose a new class\, coined simplex random features\, whi ch is even better (in fact\, optimal among a broad family). We use our new class to approximate approximation in scalable Transformers\, obtaining s ometimes substantial gains (0.5% on ImageNet) at essentially no cost.\n \n TLDR: first rigorous explanation of the efficacy of orthogonal random feat ures and derivation of a better (optimal) class.\n \n17:45 – 18:00:\n \n Speaker: Jannes Gladrow\, Cloud systems future\, Microsoft Research Cambri dge\n \nTitle: Efficient data transport over multimode light-pipes with Me gapixel images using differentiable ray tracing and Machine-learning\n \nA bstract: Retrieving images transmitted through multi-mode fibers is of gro wing interest\, thanks to their ability to confine and transport light eff iciently 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 meg apixel 8-bit spatial light modulator\, presenting data as a matrix of symb ols. Normally\, (deep-learning) decoders will incur a prohibitive O(n^2) c omputational scaling to decode n symbols in spatially scrambled data. Howe ver\, by combining a digital twin of the setup with a U-Net\, we can retri eve 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-experi ment gap by adjusting to optical imperfections. We train the pipeline end- to-end using a differentiable mutual-information estimator based on the vo n-Mises distribution which is generally applicable to phase-coding channel s. LOCATION:The small lecture theatre\, 21 Station Rd END:VEVENT END:VCALENDAR