Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
A lecture by Yuval Milo from Tel Aviv University exploring the theoretical advantages of complex number parameterizations in Structured State Space Models (SSMs), which power neural networks like S4 and Mamba. Discover formal mathematical proofs demonstrating that complex SSMs can express all mappings of real SSMs with moderate dimensions, while real SSMs require significantly higher dimensions or exponentially large parameter values. Learn about the research findings that establish clear gaps between real and complex diagonal SSMs, and how these insights extend to selectivity—a new architectural feature delivering state-of-the-art performance. The talk presents results from the paper "Provable Benefits of Complex Parameterizations for Structured State Space Models." Taking place Thursday, March 27th, 2025, at 10:30 AM in room B220, hosted by the HUJI Machine Learning Club.
