Search papers, labs, and topics across Lattice.
This paper introduces a variational approach to training a finite-temperature continuous-spin perceptron on a Gaussian mixture, allowing for a wide range of concave utilities and log-concave prior measures. By employing an interpolation method alongside log-concavity and concentration estimates, the authors derive minimax variational bounds for the limiting quenched pressure, revealing that the bounds differ primarily in the order of optimization of two variational parameters. Notably, when these optimizations commute, the bounds converge, providing a unified framework for calculating ground-state energy, training loss, and generalization error.
The variational bounds derived in this study reveal a surprising interplay between optimization order and model solution, offering a new lens on perceptron learning dynamics.
We introduce a variational approach to a finite-temperature continuous-spin perceptron trained on a Gaussian mixture. The model allows for a broad class of concave utilities and log-concave separable prior measures on the spins. By combining the interpolation method with log-concavity and concentration estimates, we derive lower and upper minimax variational bounds for the limiting quenched pressure. Remarkably, the two bounds differ only in the order of optimization of two variational parameters, while all remaining extrema are controlled by the concave--convex structure of the variational potential. Whenever the two optimizations commute, the two bounds match and identify the solution of the model. The same potential yields the fixed-point equations as stationarity conditions and provides a unified route to the computation of the ground-state energy, training loss, and generalization error.
