Théo Jules

Theo Jules, Gal Brener, Tal Kachman et al.

The training of neural networks is a complex, high-dimensional, non-convex and noisy optimization problem whose theoretical understanding is interesting both from an applicative perspective and for fundamental reasons. A core challenge is to understand the geometry and topography of the landscape that guides the optimization. In this work, we employ standard Statistical Mechanics methods, namely, phase-space exploration using Langevin dynamics, to study this landscape for an over-parameterized fully connected network performing a classification task on random data. Analyzing the fluctuation statistics, in analogy to thermal dynamics at a constant temperature, we infer a clear geometric description of the low-loss region. We find that it is a low-dimensional manifold whose dimension can be readily obtained from the fluctuations. Furthermore, this dimension is controlled by the number of data points that reside near the classification decision boundary. Importantly, we find that a quadratic approximation of the loss near the minimum is fundamentally inadequate due to the exponential nature of the decision boundary and the flatness of the low-loss region. This causes the dynamics to sample regions with higher curvature at higher temperatures, while producing quadratic-like statistics at any given temperature. We explain this behavior by a simplified loss model which is analytically tractable and reproduces the observed fluctuation statistics.

Théo Jules, Laura Michel, Adèle Douin et al.

Traditional memory writing operations proceed one bit at a time, where e.g. an individual magnetic domain is force-flipped by a localized external field. One way to increase material storage capacity would be to write several bits at a time in the bulk of the material. However, the manipulation of bits is commonly done through quasi-static operations. While simple to model, this method is known to reduce memory capacity. In this paper, we demonstrate how a reinforcement learning agent can exploit the dynamical response of a simple multi-bit mechanical system to restore its memory to full capacity. To do so, we introduce a model framework consisting of a chain of bi-stable springs, which is manipulated on one end by the external action of the agent. We show that the agent manages to learn how to reach all available states for three springs, even though some states are not reachable through adiabatic manipulation, and that both the training speed and convergence within physical parameter space are improved using transfer learning techniques. Interestingly, the agent also points to an optimal design of the system in terms of writing time. In fact, it appears to learn how to take advantage of the underlying physics: the control time exhibits a non-monotonic dependence on the internal dissipation, reaching a minimum at a cross-over shown to verify a mechanically motivated scaling relation.