Abstract | ||
---|---|---|
We approach the problem of implicit regularization in deep learning from a geometrical viewpoint. We highlight a regularization effect induced by a dynamical alignment of the neural tangent features introduced by Jacot et al. (2018), along a small number of task-relevant directions. This can be interpreted as a combined mechanism of feature selection and compression. By extrapolating a new analysis of Rademacher complexity bounds for linear models, we motivate and study a heuristic complexity measure that captures this phenomenon, in terms of sequences of tangent kernel classes along optimization paths. The code for our experiments is available as https://github.com/tfjgeorge/ntk_alignment. |
Year | Venue | Keywords |
---|---|---|
2021 | 24TH INTERNATIONAL CONFERENCE ON ARTIFICIAL INTELLIGENCE AND STATISTICS (AISTATS) | Rademacher complexity,Deep learning,Regularization (mathematics),Feature selection,Kernel (linear algebra),Tangent,Heuristic,Linear model,Algorithm,Computer science,Artificial intelligence |
DocType | Volume | ISSN |
Conference | 130 | 2640-3498 |
Citations | PageRank | References |
0 | 0.34 | 0 |
Authors | ||
7 |
Name | Order | Citations | PageRank |
---|---|---|---|
Aristide Baratin | 1 | 25 | 2.69 |
Thomas George | 2 | 4 | 1.41 |
César Laurent | 3 | 87 | 5.29 |
R Devon Hjelm | 4 | 0 | 0.34 |
Guillaume Lajoie | 5 | 8 | 6.67 |
Pascal Vincent | 6 | 7834 | 504.76 |
Simon Lacoste-Julien | 7 | 0 | 0.34 |