Two Routes to Scalable Credit Assignment without Weight Symmetry
12. července 2020
Řečníci
Daniel Kunin
Sprecher:in · 0 Follower:innen
Aran Nayebi
Sprecher:in · 0 Follower:innen
Javier Sagastuy-Brena
Sprecher:in · 0 Follower:innen
Surya Ganguli
Sprecher:in · 3 Follower:innen
Jonathan Bloom
Sprecher:in · 0 Follower:innen
Daniel Yamins
Sprecher:in · 0 Follower:innen
O prezentaci
The neural plausibility of backpropagation has long been disputed, primarily for its use of instantaneous weight transport. A variety of prior proposals that avoid weight transport fail to scale on complex tasks such as ImageNet; however, a recent proposal has reported competitive performance with backpropagation. We find that this local learning rule requires complex hyperparameter tuning that does not transfer across architectures. We identify a more robust local learning rule that transfers to deep networks on ImageNet without re-tuning hyperparameters. Nonetheless, we find a performance gap between this local rule and backpropagation that widens with increasing model depth. We formulate several non-local learning rules that address these shortcomings without requiring strict weight symmetry and match state-of-the-art performance on networks with hundreds of layers, even in the presence of noisy updates. Our results suggest two routes towards the discovery of neural implementations for credit assignment: further improvement of local rules so that they perform consistently across architectures and the identification of scalable biological implementations for our non-local learning mechanisms.The neural plausibility of backpropagation has long been disputed, primarily for its use of instantaneous weight transport. A variety of prior proposals that avoid weight transport fail to scale on complex tasks such as ImageNet; however, a recent proposal has reported competitive performance with backpropagation. We find that this local learning rule requires complex hyperparameter tuning that does not transfer across architectures. We identify a more robust local learning rule that transfers t…
Kategorie
KI und Datenwissenschaft
Kategorie · 10,8k Präsentationen
O organizátorovi (ICML 2020)
The International Conference on Machine Learning (ICML) is the premier gathering of professionals dedicated to the advancement of the branch of artificial intelligence known as machine learning. ICML is globally renowned for presenting and publishing cutting-edge research on all aspects of machine learning used in closely related areas like artificial intelligence, statistics and data science, as well as important application areas such as machine vision, computational biology, speech recognition, and robotics. ICML is one of the fastest growing artificial intelligence conferences in the world. Participants at ICML span a wide range of backgrounds, from academic and industrial researchers, to entrepreneurs and engineers, to graduate students and postdocs.
Baví vás formát? Nechte SlidesLive zachytit svou akci!
Profesionální natáčení a streamování po celém světě.
Sdílení
Doporučená videa
Prezentace na podobné téma, kategorii nebo přednášejícího
Counterfactual Transfer via Inductive Bias in Clinical Settings
Später ansehen
GCNs: From Summarization to Heterophily
Später ansehen
Flows in Probabilistic Modeling & Inference
Später ansehen
Generative Adversarial Set Transformers
Später ansehen
Evolutionary Reinforcement Learning for Sample-Efficient Multiagent Cooperation
Später ansehen
