[Paper Review] Attention-Based Structural-Plasticity

Soheil Kolouri, Nicholas Ketz+3 more|arXiv (Cornell University)|Mar 2, 2019

Domain Adaptation and Few-Shot Learning14 references4 citations

TL;DR

This paper proposes an attention-based selective plasticity mechanism inspired by the brain's cholinergic system to combat catastrophic forgetting in continual learning. By learning synaptic importance via Hebbian rule in parallel with backpropagation, the method dynamically protects critical weights, achieving performance on par with state-of-the-art methods like EWC and Synaptic Intelligence on Permuted and Split MNIST without extensive hyperparameter tuning.

ABSTRACT

Catastrophic forgetting/interference is a critical problem for lifelong learning machines, which impedes the agents from maintaining their previously learned knowledge while learning new tasks. Neural networks, in particular, suffer plenty from the catastrophic forgetting phenomenon. Recently there has been several efforts towards overcoming catastrophic forgetting in neural networks. Here, we propose a biologically inspired method toward overcoming catastrophic forgetting. Specifically, we define an attention-based selective plasticity of synapses based on the cholinergic neuromodulatory system in the brain. We define synaptic importance parameters in addition to synaptic weights and then use Hebbian learning in parallel with backpropagation algorithm to learn synaptic importances in an online and seamless manner. We test our proposed method on benchmark tasks including the Permuted MNIST and the Split MNIST problems and show competitive performance compared to the state-of-the-art methods.

Motivation & Objective

• To address catastrophic forgetting in lifelong learning, where neural networks overwrite previously learned knowledge when exposed to new tasks.
• To develop a biologically plausible mechanism that selectively protects important synapses during continual learning.
• To integrate top-down attention signals with synaptic plasticity using Hebbian learning for online, dynamic weight protection.
• To demonstrate competitive performance against state-of-the-art continual learning methods on benchmark tasks.

Proposed method

• Introduces a synaptic importance parameter alongside traditional weights, updated via a Hebbian learning rule based on pre- and post-synaptic importance signals.
• Employs a contrastive excitatory backpropagation (c-EB) mechanism to generate top-down attention maps using predicted class labels.
• Uses the c-EB top-down signal to modulate synaptic importance, emphasizing task-relevant features and suppressing distractions.
• Updates synaptic importance in parallel with standard backpropagation, enabling online, seamless learning of importance without additional training phases.
• Applies Oja’s learning rule to update synaptic importance based on pre- and post-synaptic importance, ensuring stability and biological plausibility.
• Leverages the cholinergic neuromodulatory system as a biological inspiration for attention-driven plasticity modulation.

Experimental results

Research questions

• RQ1 Can a biologically inspired attention mechanism improve lifelong learning by selectively protecting important synapses?
• RQ2 How does attention-based synaptic importance learning compare to existing regularization-based continual learning methods?
• RQ3 Can Hebbian learning of synaptic importance be effectively combined with backpropagation in an online, end-to-end manner?
• RQ4 Does the proposed method maintain high performance on previously learned tasks while learning new ones, without extensive hyperparameter tuning?

Key findings

• The proposed method achieves performance on par with state-of-the-art methods such as EWC and Synaptic Intelligence on the Permuted MNIST benchmark.
• On the Split MNIST task, the method maintains high accuracy across all five sequential digit-pair tasks, outperforming a vanilla network and matching the performance of Synaptic Intelligence.
• The method shows robustness to distribution shifts and non-stationary data, effectively mitigating catastrophic forgetting in sequential learning scenarios.
• Despite minimal hyperparameter tuning, the method achieves competitive results, suggesting strong generalization and stability of the attention-based plasticity mechanism.
• Attention maps generated via c-EB successfully highlight task-relevant features, demonstrating the model's ability to focus on important inputs during learning.