Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

12 mJ per Class On-Device Online Few-Shot Class-Incremental Learning

Yoga Esa Wibowo, Cristian Cioflan, Thorir Mar Ingolfsson, Michael Hersche, Leo Zhao, Abbas Rahimi, Luca Benini

TL;DR

Online Few-Shot Class-Incremental Learning is introduced, based on a lightweight model consisting of a pre-trained and metalearned feature extractor and an expandable explicit memory storing the class prototypes, which allows learning previously unseen classes based on only a few examples with one single pass.

Abstract

Few-Shot Class-Incremental Learning (FSCIL) enables machine learning systems to expand their inference capabilities to new classes using only a few labeled examples, without forgetting the previously learned classes. Classical backpropagation-based learning and its variants are often unsuitable for battery-powered, memory-constrained systems at the extreme edge. In this work, we introduce Online Few-Shot Class-Incremental Learning (O-FSCIL), based on a lightweight model consisting of a pretrained and metalearned feature extractor and an expandable explicit memory storing the class prototypes. The architecture is pretrained with a novel feature orthogonality regularization and metalearned with a multi-margin loss. For learning a new class, our approach extends the explicit memory with novel class prototypes, while the remaining architecture is kept frozen. This allows learning previously unseen classes based on only a few examples with one single pass (hence online). O-FSCIL obtains an average accuracy of 68.62% on the FSCIL CIFAR100 benchmark, achieving state-of-the-art results. Tailored for ultra-low-power platforms, we implement O-FSCIL on the 60 mW GAP9 microcontroller, demonstrating online learning capabilities within just 12 mJ per new class.

12 mJ per Class On-Device Online Few-Shot Class-Incremental Learning

TL;DR

Online Few-Shot Class-Incremental Learning is introduced, based on a lightweight model consisting of a pre-trained and metalearned feature extractor and an expandable explicit memory storing the class prototypes, which allows learning previously unseen classes based on only a few examples with one single pass.

Abstract

Few-Shot Class-Incremental Learning (FSCIL) enables machine learning systems to expand their inference capabilities to new classes using only a few labeled examples, without forgetting the previously learned classes. Classical backpropagation-based learning and its variants are often unsuitable for battery-powered, memory-constrained systems at the extreme edge. In this work, we introduce Online Few-Shot Class-Incremental Learning (O-FSCIL), based on a lightweight model consisting of a pretrained and metalearned feature extractor and an expandable explicit memory storing the class prototypes. The architecture is pretrained with a novel feature orthogonality regularization and metalearned with a multi-margin loss. For learning a new class, our approach extends the explicit memory with novel class prototypes, while the remaining architecture is kept frozen. This allows learning previously unseen classes based on only a few examples with one single pass (hence online). O-FSCIL obtains an average accuracy of 68.62% on the FSCIL CIFAR100 benchmark, achieving state-of-the-art results. Tailored for ultra-low-power platforms, we implement O-FSCIL on the 60 mW GAP9 microcontroller, demonstrating online learning capabilities within just 12 mJ per new class.
Paper Structure (18 sections, 4 equations, 3 figures, 4 tables)

This paper contains 18 sections, 4 equations, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Few-shot Class-incremental Learning
  4. On-device continual learning
  5. task description
  6. Online Few-shot Class-incremental Learning
  7. Light-weight backbone
  8. Pretraining
  9. Metalearning
  10. Hardware deployment
  11. Deployment
  12. On-Device learning
  13. Experimental results
  14. CIFAR100 benchmark
  15. Ablation study
  16. ...and 3 more sections

Figures (3)

  • Figure 1: Inference (a), on-device learning a new class (b), and server-side metalearning (c) modes of O-FSCIL. Modules colored in orange are updated, while grey ones are frozen. During pretraining, we replaced the prototype computation and EM update from (b) with an FCR-like FCC classifier, with all three sections jointly trained.
  • Figure 2: Average number of operations per cycle given the number of active cores, for backbone inference (left), inference (centre), and backpropagation update (right).
  • Figure 3: The representation precision in the episodic memory impacts the accuracy of MobileNetV2_x4-based model, considering 100 class prototypes stored in the memory.