Dataset Information

Random synaptic feedback weights support error backpropagation for deep learning.

ABSTRACT: The brain processes information through multiple layers of neurons. This deep architecture is representationally powerful, but complicates learning because it is difficult to identify the responsible neurons when a mistake is made. In machine learning, the backpropagation algorithm assigns blame by multiplying error signals with all the synaptic weights on each neuron's axon and further downstream. However, this involves a precise, symmetric backward connectivity pattern, which is thought to be impossible in the brain. Here we demonstrate that this strong architectural constraint is not required for effective error propagation. We present a surprisingly simple mechanism that assigns blame by multiplying errors by even random synaptic weights. This mechanism can transmit teaching signals across multiple layers of neurons and performs as effectively as backpropagation on a variety of tasks. Our results help reopen questions about how the brain could use error signals and dispel long-held assumptions about algorithmic constraints on learning.

SUBMITTER: Lillicrap TP

PROVIDER: S-EPMC5105169 | biostudies-literature | 2016 Nov

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Random synaptic feedback weights support error backpropagation for deep learning.

Lillicrap Timothy P TP Cownden Daniel D Tweed Douglas B DB Akerman Colin J CJ

Nature communications 20161108

The brain processes information through multiple layers of neurons. This deep architecture is representationally powerful, but complicates learning because it is difficult to identify the responsible neurons when a mistake is made. In machine learning, the backpropagation algorithm assigns blame by multiplying error signals with all the synaptic weights on each neuron's axon and further downstream. However, this involves a precise, symmetric backward connectivity pattern, which is thought to be ...[more]

PMID: 27824044

Similar Datasets

Project description:Humans monitor their behavior to optimize performance, which presumably relies on stable representations of correct responses. During second language (L2) learning, however, stable representations have yet to be formed while knowledge of the first language (L1) can interfere with learning, which in some cases results in persistent errors. In order to examine how correct L2 representations are stabilized, this study examined performance monitoring in the learning process of second language learners for a feature that conflicts with their first language. Using EEG, we investigated if L2 learners in a feedback-guided word gender assignment task showed signs of error detection in the form of an error-related negativity (ERN) before and after receiving feedback, and how feedback is processed. The results indicated that initially, response-locked negativities for correct (CRN) and incorrect (ERN) responses were of similar size, showing a lack of internal error detection when L2 representations are unstable. As behavioral performance improved following feedback, the ERN became larger than the CRN, pointing to the first signs of successful error detection. Additionally, we observed a second negativity following the ERN/CRN components, the amplitude of which followed a similar pattern as the previous negativities. Feedback-locked data indicated robust FRN and P300 effects in response to negative feedback across different rounds, demonstrating that feedback remained important in order to update memory representations during learning. We thus show that initially, L2 representations may often not be stable enough to warrant successful error monitoring, but can be stabilized through repeated feedback, which means that the brain is able to overcome L1 interference, and can learn to detect errors internally after a short training session. The results contribute a different perspective to the discussion on changes in ERN and FRN components in relation to learning, by extending the investigation of these effects to the language learning domain. Furthermore, these findings provide a further characterization of the online learning process of L2 learners.

Dataset Information

Random synaptic feedback weights support error backpropagation for deep learning.

Publications

Random synaptic feedback weights support error backpropagation for deep learning.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets