We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations

DBLP - CS Bibliography


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: A contrastive rule for meta-learning

Abstract: Meta-learning algorithms leverage regularities that are present on a set of tasks to speed up and improve the performance of a subsidiary learning process. Recent work on deep neural networks has shown that prior gradient-based learning of meta-parameters can greatly improve the efficiency of subsequent learning. Here, we present a biologically plausible meta-learning algorithm based on equilibrium propagation. Instead of explicitly differentiating the learning process, our contrastive meta-learning rule estimates meta-parameter gradients by executing the subsidiary process more than once. This avoids reversing the learning dynamics in time and computing second-order derivatives. In spite of this, and unlike previous first-order methods, our rule recovers an arbitrarily accurate meta-parameter update given enough compute. We establish theoretical bounds on its performance and present experiments on a set of standard benchmarks and neural network architectures.
Comments: 29 pages, 9 figures
Subjects: Machine Learning (cs.LG); Neural and Evolutionary Computing (cs.NE); Neurons and Cognition (q-bio.NC)
Cite as: arXiv:2104.01677 [cs.LG]
  (or arXiv:2104.01677v2 [cs.LG] for this version)

Submission history

From: João Sacramento [view email]
[v1] Sun, 4 Apr 2021 19:45:41 GMT (639kb,D)
[v2] Mon, 19 Apr 2021 21:04:58 GMT (713kb,D)

Link back to: arXiv, form interface, contact.