A rapid and efficient learning rule for biological neural circuits

Eren Sezener, Agnieszka Grabska‐Barwińska, Dimitar Kostadinov, Maxime Beau, Sanjukta Krishnagopal, David Budden, Marcus Hütter, Joel Veness, Matthew Botvinick, Claudia Clopath, Michael Häusser, Peter E. LathambioRxiv 2021 · DOI

TL;DR

This paper introduces Dendritic Gated Networks (DGNs), a novel biologically plausible learning algorithm that mimics dendritic inhibition for efficient learning and resistance to catastrophic forgetting. DGNs outperform traditional backpropagation on several tasks and show promising structural and functional parallels with cerebellar circuits, supported by in vivo experimental evidence.

What Problem Does This Paper Address?

The dominant paradigm in deep learning, backpropagation, is highly effective but suffers from two major drawbacks: it's biologically implausible, and it leads to "catastrophic forgetting" where networks forget old tasks when learning new ones. Neuroscience suggests that biological learning is more nuanced, involving precise control over synaptic changes. However, the exact mechanisms by which the brain determines which synapses to modify and by how much remain unclear. This paper aims to bridge this gap by proposing a learning rule that is both biologically plausible and addresses the limitations of current artificial neural networks, particularly their susceptibility to forgetting and data inefficiency.

Methods

The core innovation is the Dendritic Gated Network (DGN). Unlike standard networks where neurons have a single set of weights, DGNs feature dendritic branches, each with its own set of weights. These branches are gated on and off by inhibitory interneurons, controlled by external input. This means a neuron's overall activity is a sum of contributions from its active dendritic branches.

Key aspects of the DGN method include:

  • Dendritic Gating: Interneurons selectively inhibit specific dendritic branches, effectively controlling which inputs contribute to the neuron's output. This gating is implemented via a random, fixed linear transformation of the input, not learned.
  • Local Learning Rule: DGNs use a gated version of the delta rule, a biologically plausible learning mechanism that only requires local information (presynaptic activity and target output). This contrasts with backpropagation's need for global error signals and knowledge of outgoing weights.
  • Biological Plausibility: The architecture and learning rule are designed to mimic biological processes, particularly the role of interneurons in shaping neuronal responses and the cerebellum's structure.
  • Efficiency and Robustness: DGNs are designed to be data-efficient and resistant to catastrophic forgetting without requiring explicit task boundary information.

Datasets

The paper evaluates DGNs on the following datasets/benchmarks:

  • Permuted MNIST: A benchmark for continual learning, where the network must learn sequential tasks involving randomly permuted MNIST digits.
  • SARCOS benchmark: An inverse kinematics dataset from a 7 degree-of-freedom robot arm.
  • Vestibulo-ocular reflex (VOR) adaptation task: A simulation of how animals adapt eye movements to changes in visual feedback.

Key Insights and Ideas

  • Dendritic Inhibition as a Learning Mechanism: The paper proposes that targeted inhibition of dendritic branches by interneurons is a key mechanism for efficient and selective learning, allowing networks to learn new tasks without overwriting old ones.
  • Random Gating for Efficiency: Surprisingly, the random, unlearned nature of the dendritic gating mechanism is crucial. It allows DGNs to efficiently represent complex, arbitrary nonlinear functions, making them robust to forgetting.
  • Cerebellar Parallels: The DGN architecture strongly resembles the cerebellum, with Purkinje cells acting as hidden layers and cerebellar nuclear neurons as output layers, and molecular layer interneurons performing the gating function. This suggests a potential biological implementation of the DGN algorithm.
  • Experimental Validation of Dendritic Gating: In vivo two-photon calcium imaging provides direct experimental evidence that molecular layer interneurons can indeed suppress calcium signals in individual dendritic branches of Purkinje cells, supporting a key prediction of the DGN model.

Key Results

  • Catastrophic Forgetting Resistance: DGNs show significantly less forgetting on the permuted MNIST task compared to standard MLPs, performing comparably to more complex methods like Elastic Weight Consolidation (EWC) without requiring task boundary information.
  • Data Efficiency: DGNs are significantly more data-efficient than conventional artificial networks.
  • Performance: DGNs perform well on various tasks, sometimes exceeding backpropagation. For example, on the SARCOS inverse kinematics task, a DGN with 5000 branches achieved the best result among compared methods, outperforming them by a factor of 50.
  • Cerebellar Modeling: The DGN model successfully replicates VOR adaptation and inverse kinematics tasks, aligning with the cerebellum's known functions.
  • In Vivo Evidence: Experimental results demonstrate that MLIs can locally inhibit Purkinje cell dendritic branches, providing strong biological support for the DGN's gating mechanism.

Limitations and Future Work

  • Binary Gating: The current DGN model uses binary (on/off) gating, which is challenging to implement biologically. The authors note that softer gates, where the amount of gating is a function of input, are more likely in the brain and found that simulations with soft gates yielded virtually identical results, suggesting flexibility.
  • MLI Input Dependence: A key prediction of the DGN that awaits experimental validation is that MLI activity should depend predominantly on parallel fiber input, and their input-output relationship should change slowly relative to Purkinje cell learning timescales.
  • Generalizability to Other Brain Circuits: While the DGN is well-suited for the cerebellum, its application to other brain circuits like the neocortex might require modifications due to differences in neuronal activity (e.g., non-linearities).
  • Future Work: The authors suggest that DGNs could be widely useful in machine learning and inspire further investigation into biological learning mechanisms in various brain circuits. They also propose further experiments to test the stability of MLI input-output relationships and to map parallel fiber connectivity patterns in tasks like VOR learning to distinguish between DGN-like and MLP-like implementations.

Summary updated

Abstract

Abstract The dominant view in neuroscience is that changes in synaptic weights underlie learning. It is unclear, however, how the brain is able to determine which synapses should change, and by how much. This uncertainty stands in sharp contrast to deep learning, where changes in weights are explicitly engineered to optimize performance. However, the main tool for that, backpropagation, has two problems. One is neuro-science related: it is not biologically plausible. The other is inherent: networks trained with this rule tend to forget old tasks when learning new ones. Here we introduce the Dendritic Gated Network (DGN), a variant of the Gated Linear Network, which offers a biologically plausible alternative to backpropagation. DGNs combine dendritic ‘gating’ (whereby interneurons target dendrites to shape neuronal responses) with local learning rules to yield provably efficient performance. They are significantly more data efficient than conventional artificial networks, and are highly resistant to forgetting. Consequently, they perform well on a variety of tasks, in some cases better than backpropagation. Importantly, DGNs have structural and functional similarities to the cerebellum, a link that we strengthen by using in vivo two-photon calcium imaging to show that single interneurons suppress activity in individual dendritic branches of Purkinje cells, a key feature of the model. Thus, DGNs leverage targeted dendritic inhibition and local learning – two features ubiquitous in the brain – to achieve fast and efficient learning.
BibTeX
@misc{sezener2021rapid,
  author = {Eren Sezener and Agnieszka Grabska‐Barwińska and Dimitar Kostadinov and Maxime Beau and Sanjukta Krishnagopal and David Budden and Marcus Hütter and Joel Veness and Matthew Botvinick and Claudia Clopath and Michael Häusser and Peter E. Latham},
  title = {{A rapid and efficient learning rule for biological neural circuits}},
  year = {2021},
  doi = {10.1101/2021.03.10.434756},
  howpublished = {bioRxiv},
}
Figure