[Theory] [Theory Lunch] REMINDER: Lorenzo Orecchia, Wednesday 2/28 12:30-1:30pm, JCL 390

[Gabe Schoenbach]

Hi all — please join us *today at 12:30pm* for another theory lunch!
Details below:

*****
*Date: *February 28, 2024
*Time: *12:30 CT
*Location: *JCL 390

*Title: *Deep Learning meets SDP Rounding: Improved Training of Binary
Neural Networks

*Speaker: *Lorenzo Orecchia (UChicago)

*Abstract:* Binary Neural Networks (BNNs) are neural networks whose weights
and activations are restricted to be binary, i.e., {+1,-1}. BNNs promise to
greatly improve the deployability of deep learning by requiring simpler
hardware and consuming much less energy, while achieving comparable
performance to general floating-point neural networks. While much effort
has been devoted to design and optimize BNN network architectures, less
effort has been devoted to the crucial problem of training BNNs. This
problem adds to the non-convexity of neural network training the additional
challenge of a combinatorial space of solutions. We attack this problem by
leveraging a natural non-convex SDP formulation and design efficient
heuristics for minimizing it via a gradient flow over an appropriately
chosen manifold. Surprisingly, in our experiments, this theoretical
approach immediately yields results that improve on the state-of-the-art
for BNN training.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cs.uchicago.edu/pipermail/theory/attachments/20240228/f1d227f9/attachment.html>