Mesoscopic description
A theoretical approach to understanding neural network learning dynamics by characterizing the statistical distribution of overlaps rather than individual overlaps, applicable to wide networks in which the number of overlaps diverges.
Timeline (last 90 days)
The paper proposes a mesoscopic description that understands the learning process by characterizing the distribution of overlaps rather than tracking individual overlaps.
The study's methodology may generalize to deeper networks, but the paper itself focuses only on two-layer networks.
All Facts (2)
The paper proposes a mesoscopic description that understands the learning process by characterizing the distribution of overlaps rather than tracking individual overlaps.
50%UnverifiedThe study's methodology may generalize to deeper networks, but the paper itself focuses only on two-layer networks.
50%