[Paper Review] FearNet: Brain-Inspired Model for Incremental Learning

Ronald Kemker, Christopher Kanan|arXiv (Cornell University)|Nov 28, 2017

Memory and Neural Mechanisms16 references271 citations

TL;DR

FearNet is a brain-inspired incremental learning model with dual-memory (HC for recent memories and mPFC for long-term storage) plus a BLA module to select memory source, using pseudorehearsal to consolidate memories. It achieves state-of-the-art results on CIFAR-100, CUB-200, and AudioSet with a small memory footprint.

ABSTRACT

Incremental class learning involves sequentially learning classes in bursts of examples from the same class. This violates the assumptions that underlie methods for training standard deep neural networks, and will cause them to suffer from catastrophic forgetting. Arguably, the best method for incremental class learning is iCaRL, but it requires storing training examples for each class, making it challenging to scale. Here, we propose FearNet for incremental class learning. FearNet is a generative model that does not store previous examples, making it memory efficient. FearNet uses a brain-inspired dual-memory system in which new memories are consolidated from a network for recent memories inspired by the mammalian hippocampal complex to a network for long-term storage inspired by medial prefrontal cortex. Memory consolidation is inspired by mechanisms that occur during sleep. FearNet also uses a module inspired by the basolateral amygdala for determining which memory system to use for recall. FearNet achieves state-of-the-art performance at incremental class learning on image (CIFAR-100, CUB-200) and audio classification (AudioSet) benchmarks.

Motivation & Objective

• Motivate incremental class learning and address catastrophic forgetting without storing all past data.
• Propose FearNet with a hippocampal-inspired recent memory and a medial prefrontal cortex-inspired long-term memory system.
• Incorporate a basolateral amygdala-inspired selector to choose memory for recall.
• Use pseudorehearsal via a generative autoencoder to consolidate recent memories into long-term storage.
• Demonstrate scalability and memory efficiency on image and audio benchmarks.

Proposed method

• Three neural modules: HC for recent memories, mPFC for long-term storage via a reconstructive autoencoder, and BLA to decide recall source.
• Pseudorehearsal through a generative autoencoder to create pseudo-examples from class statistics (means and covariances) for memory consolidation without storing past data.
• HC stores exemplars per class; after consolidation, exemplars are moved to mPFC and HC is cleared.
• mPFC is trained with a loss combining supervised classification and reconstruction (L_class + L_recon).
• BLA outputs a probability A(x) guiding whether to recall from HC or mPFC; final prediction blends HC and mPFC posteriors with a heuristic.

Experimental results

Research questions

• RQ1 Can FearNet achieve superior incremental learning performance without storing past training data?
• RQ2 How does a brain-inspired dual-memory architecture with pseudorehearsal compare to existing methods on large-scale benchmarks?
• RQ3 What is the role of memory consolidation timing (sleep) on base-knowledge retention and new-class recall?
• RQ4 How memory footprint and modality (images, audio) affect incremental learning performance?
• RQ5 Can a memory selector (BLA) effectively route recall to the appropriate memory system?

Key findings

• FearNet achieves state-of-the-art Omega_base and Omega_all across CIFAR-100, CUB-200, and AudioSet compared to baseline incremental methods.
• FearNet’s performance closely tracks offline MLP baselines, especially for base-knowledge and overall retention.
• Pseudorehearsal via mPFC reconstruction enables memory consolidation from HC to long-term storage without storing prior data.
• BLA-based memory selection effectively distinguishes when to use HC versus mPFC for recall, with near-oracle performance in many cases.
• FearNet maintains a relatively small memory footprint by storing class statistics (means and covariances) rather than raw data.
• FearNet remains robust across multi-modal incremental tasks and shows scalability to 100–200 classes.