Special Seminars

Bio for Luis Serrano:

Luis Serrano did his undergraduate and masters in math at the University of Waterloo, his PhD in algebraic combinatorics at the University of Michigan, and a postdoctoral fellowship at the University of Quebec at Montreal. He then went to industry, working in AI at Google, Apple, in large language models at Cohere, and in quantum AI at Zapata Computing. He has taught courses in AI at platforms like Udacity and Coursera, and hosts the popular YouTube channel Serrano Academy with over 180 thousand subscribers. He’s also the author of the Amazon bestseller Grokking Machine Learning.

As part of a visit to Perimeter of a delegation from the ELI Beamlines laser facility in the Czech Republic, Dr. Bulanov will speak about potential topics for collaboration between Perimeter Institute and 
ELI theorists on topics related to high energy laser physics. To highlight the interplay between theory and experiment, Dr. Bulanov will briefly mention two experiments where this was realized and is currently planned at the ELI Beamlines facility.

High school students learn how to express the solution of a quadratic equation in one unknown in terms of its three coefficients.  Why does this formula matter? We offer an answer in terms of discriminants and data. This lecture invites the audience to a journey towards non-linear algebra.

We explore the interplay of quantum computing and machine learning to advance experimental protocols for observing measurement-induced phase transitions (MIPT) in quantum devices. In particular, we focus on trapped ion monitored circuits and apply the cross entropy benchmark recently introduced by [Li et al., Phys. Rev. Lett. 130, 220404 (2023)], which can mitigate the postselection problem. By doing so, we reduce the number of projective measurements -- the sample complexity required per random circuit realization, which is a critical limiting resource in real devices. Since these projective measurement outcomes form a classical probability distribution, they are suitable for learning with a standard machine learning generative model. In this work, we use a recurrent neural network (RNN) to learn a representation of the measurement record for a native trapped-ion MIPT, and show that using this generative model can substantially reduce the number of measurements required to accurately estimate the cross entropy. This illustrates the potential of combining quantum computing and machine learning to overcome practical challenges in realizing quantum experiments. 

Rydberg atom arrays have emerged as powerful quantum simulators, capable of preparing strongly correlated phases of matter that are potentially challenging to access with classical computational methods. A major focus has been on realizing these arrays on frustrated geometries, aiming to stabilize exotic many-body states like spin liquids. In this talk, I will show how two-dimensional recurrent neural network (RNN) wave functions can be used to study the ground states of Rydberg atom arrays on the kagome lattice. For Hamiltonians previously investigated in this geometry, I will demonstrate that the RNN finds no evidence for exotic spin liquid phases or emergent glassiness. In particular, I will argue that signals of glassy behavior, such as a nonzero Edwards-Anderson order parameter seen in quantum Monte Carlo (QMC) studies, may arise from artifacts related to long autocorrelation times. These results highlight the potential of language model-inspired approaches, like RNNs, for advancing the study of frustrated quantum systems and Rydberg atom physics more broadly.

arXiv paper: https://arxiv.org/pdf/2405.20384

A large amount of effort has recently been put into understanding the barren plateau phenomenon. In this perspective talk, we face the increasingly loud elephant in the room and ask a question that has been hinted at by many but not explicitly addressed: Can the structure that allows one to avoid barren plateaus also be leveraged to efficiently simulate the loss classically? We present a case-by-case argument that commonly used models with provable absence of barren plateaus are also in a sense classically simulable, provided that one can collect some classical data from quantum devices during an initial data acquisition phase. This follows from the observation that barren plateaus result from a curse of dimensionality, and that current approaches for solving them end up encoding the problem into some small, classically simulable, subspaces. We end by discussing caveats in our arguments including the limitations of average case arguments, the role of smart initializations, models that fall outside our assumptions, the potential for provably superpolynomial advantages and the possibility that, once larger devices become available, parametrized quantum circuits could heuristically outperform our analytic expectations.