[Paper Review] Meta-Learned Attribute Self-Gating for Continual Generalized Zero-Shot Learning

Vinay Kumar Verma, Kevin Liang+2 more|arXiv (Cornell University)|Feb 23, 2021

Domain Adaptation and Few-Shot Learning4 citations

TL;DR

This paper proposes Meta-Continual Zero-Shot Learning (MCZSL), a fast, scalable approach that uses meta-learning with self-gated attributes and scaled class normalization to achieve state-of-the-art performance in generalized zero-shot learning (GZSL) and continual GZSL, outperforming generative-model-based methods by over 100× faster training while mitigating catastrophic forgetting in sequential learning settings.

ABSTRACT

Zero-shot learning (ZSL) has been shown to be a promising approach to generalizing a model to categories unseen during training by leveraging class attributes, but challenges still remain. Recently, methods using generative models to combat bias towards classes seen during training have pushed the state of the art of ZSL, but these generative models can be slow or computationally expensive to train. Additionally, while many previous ZSL methods assume a one-time adaptation to unseen classes, in reality, the world is always changing, necessitating a constant adjustment for deployed models. Models unprepared to handle a sequential stream of data are likely to experience catastrophic forgetting. We propose a meta-continual zero-shot learning (MCZSL) approach to address both these issues. In particular, by pairing self-gating of attributes and scaled class normalization with meta-learning based training, we are able to outperform state-of-the-art results while being able to train our models substantially faster ($>100\ imes$) than expensive generative-based approaches. We demonstrate this by performing experiments on five standard ZSL datasets (CUB, aPY, AWA1, AWA2 and SUN) in both generalized zero-shot learning and generalized continual zero-shot learning settings.

Motivation & Objective

• To address the limitations of existing zero-shot learning (ZSL) methods that rely on expensive generative models for unseen class generalization.
• To enable continual zero-shot learning where models adapt to new classes sequentially without catastrophic forgetting.
• To improve robustness and efficiency by replacing costly generative training with self-gating and normalization mechanisms.
• To develop a meta-learning framework that enables fast adaptation to few-shot, unseen classes in dynamic environments.
• To demonstrate state-of-the-art performance in both generalized ZSL and continual GZSL across five standard benchmarks.

Proposed method

• Introduces self-gating on class attribute embeddings to suppress noisy or irrelevant attribute dimensions and learn robust, class-specific representations.
• Applies scaled class normalization (SCN) to mitigate bias toward seen classes and improve zero-shot generalization.
• Employs first-order meta-learning (e.g., MAML) to enable fast adaptation from few samples during continual learning.
• Uses a fixed-size reservoir to store a subset of past task samples, enabling continual learning without full replay.
• Combines meta-learning with attribute self-gating and SCN in an end-to-end trainable model for joint optimization.
• Trains the model using a few-shot, task-based meta-learning objective, where each task simulates a new class distribution.

Experimental results

Research questions

• RQ1 Can self-gating on attributes improve robustness and generalization in zero-shot learning without relying on generative models?
• RQ2 Can scaled class normalization effectively reduce bias toward seen classes in generalized zero-shot learning?
• RQ3 Can meta-learning enable fast adaptation to new, unseen classes with only a few samples in a continual learning setup?
• RQ4 How does the proposed MCZSL model perform in continual GZSL compared to strong baselines across multiple datasets?
• RQ5 What is the contribution of each component (self-gating, SCN, meta-learning) to overall performance in continual and generalized ZSL?

Key findings

• MCZSL achieves state-of-the-art performance on five standard GZSL datasets (CUB, aPY, AWA1, AWA2, SUN), outperforming prior generative-model-based methods.
• The model trains over 100× faster than generative baselines while maintaining or exceeding their accuracy.
• In fixed continual GZSL, MCZSL achieves a harmonic mean of 53.12 on AWA1 and 51.31 on CUB, with ablation showing a 3.23% drop when meta-learning is removed.
• In dynamic continual GZSL, MCZSL achieves absolute gains of 10.92%, 7.01%, 4.80%, and 2.30% on CUB, AWA1, AWA2, and SUN, respectively.
• Ablation studies confirm that self-gating improves harmonic mean by 1.02% on CUB and 2.04% on AWA1, while removing class normalization reduces performance by 5.91% on AWA1.
• The reservoir size of 14×number of classes per task yields optimal performance, with consistent gains across fixed and dynamic continual learning protocols.