Machine Learning / Artificial Intelligence Seminar - Veit Elser
November 13, 2026 10:00AM—11:00AM
Location:
In Person
-
Newell-Simon 3305
Speaker:
VEIT ELSER,
Professor, Department of Physics, Cornell University
https://physics.cornell.edu/veit-elser
The public mainly interacts with data centers through symbols. For some, the fact that most of the processing involves floating point multiplications will come as a surprise. But for most researchers in AI, the idea of representing information continuously is so ingrained it is never questioned. They will not deny it would be nice if neurons produced discrete outputs, but training such networks is not possible with the gradient-based method that has been in place for the past 40 years. In this talk I describe a method for training networks ofBoolean threshold functions. Instead of computing gradients, the optimizer projects variables to two kinds of constraints. One projection acts on each neuron in isolation, bringing consistency between its inputs, outputs and weights. The second constraint implements the network's connectivity, and the projection equates one neuron's output to another neuron's input. The two projections are applied iteratively as described in my ACO talk "A greedy method for solving even hard problems". The method has not yet been applied to LLMs, but results on some challenging data sets are very encouraging. It's a technology worth investigating for the simple fact that inference requires no multiplications when the network's neurons only output +1 or -1.
—
Veit Elser is a professor of physics at Cornell. In high school he was undecided between math and physics, but ended up with physics degrees, first at Caltech and then UC Berkeley. His PhD thesis (1984) on various topics in statistical mechanics was unusual because it was typeset in TeX and carried out without an advisor. He then worked at AT&T Bell Laboratories, again doing physics. By a lucky accident, the 3D Penrose tiles he built at Berkeley turned out to be a good model of some new materials called quasicrystals. In 1988 he moved to Cornell, where his physics research took an abrupt turn in 2001 when he learned about the "phase problem" and the strange algorithms being used to solve it.
Faculty Host: Nihar Shah
Note ⇢ Veit will also be giving an ACO Seminar on November 12, 3:00 pm, on A greedy method for solving even hard problems.
For More Information:
nihars@cs.cmu.edu