Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

EMN: Brain-inspired Elastic Memory Network for Quick Domain Adaptive Feature Mapping

Jianming Lv, Chengjun Wang, Depin Liang, Qianli Ma, Wei Chen, Xueqi Cheng

TL;DR

The paper tackles efficient domain adaptation under unlabeled target data by freezing the feature extractor and learning a fast mapping via a brain-inspired Elastic Memory Network (EMN). EMN performs impulse-based memory encoding, stores associations as distributed Gaussian memories, retrieves decisions through confidence-weighted fusion, and applies reinforced memorization with pseudo labels to adapt quickly to new domains. Across four real-world cross-domain datasets, EMN delivers up to 10% accuracy gains at a fraction of the time cost of traditional gradient-based domain adaptation methods, demonstrating strong edge-device potential. The work introduces the DAMap formulation and demonstrates how distributed memories can enable robust, low-cost domain adaptation without heavy backpropagation.

Abstract

Utilizing unlabeled data in the target domain to perform continuous optimization is critical to enhance the generalization ability of neural networks. Most domain adaptation methods focus on time-consuming optimization of deep feature extractors, which limits the deployment on lightweight edge devices. Inspired by the memory mechanism and powerful generalization ability of biological neural networks in human brains, we propose a novel gradient-free Elastic Memory Network, namely EMN, to support quick fine-tuning of the mapping between features and prediction without heavy optimization of deep features. In particular, EMN adopts randomly connected neurons to memorize the association of features and labels, where the signals in the network are propagated as impulses, and the prediction is made by associating the memories stored on neurons based on their confidence. More importantly, EMN supports reinforced memorization of feature mapping based on unlabeled data to quickly adapt to a new domain. Experiments based on four cross-domain real-world datasets show that EMN can achieve up to 10% enhancement of performance while only needing less than 1% timing cost of traditional domain adaptation methods.

EMN: Brain-inspired Elastic Memory Network for Quick Domain Adaptive Feature Mapping

TL;DR

The paper tackles efficient domain adaptation under unlabeled target data by freezing the feature extractor and learning a fast mapping via a brain-inspired Elastic Memory Network (EMN). EMN performs impulse-based memory encoding, stores associations as distributed Gaussian memories, retrieves decisions through confidence-weighted fusion, and applies reinforced memorization with pseudo labels to adapt quickly to new domains. Across four real-world cross-domain datasets, EMN delivers up to 10% accuracy gains at a fraction of the time cost of traditional gradient-based domain adaptation methods, demonstrating strong edge-device potential. The work introduces the DAMap formulation and demonstrates how distributed memories can enable robust, low-cost domain adaptation without heavy backpropagation.

Abstract

Utilizing unlabeled data in the target domain to perform continuous optimization is critical to enhance the generalization ability of neural networks. Most domain adaptation methods focus on time-consuming optimization of deep feature extractors, which limits the deployment on lightweight edge devices. Inspired by the memory mechanism and powerful generalization ability of biological neural networks in human brains, we propose a novel gradient-free Elastic Memory Network, namely EMN, to support quick fine-tuning of the mapping between features and prediction without heavy optimization of deep features. In particular, EMN adopts randomly connected neurons to memorize the association of features and labels, where the signals in the network are propagated as impulses, and the prediction is made by associating the memories stored on neurons based on their confidence. More importantly, EMN supports reinforced memorization of feature mapping based on unlabeled data to quickly adapt to a new domain. Experiments based on four cross-domain real-world datasets show that EMN can achieve up to 10% enhancement of performance while only needing less than 1% timing cost of traditional domain adaptation methods.
Paper Structure (20 sections, 17 equations, 15 figures, 7 tables)

This paper contains 20 sections, 17 equations, 15 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related works
  3. Unsupervised Domain Adaptation (UDA)
  4. ANNs without Gradient Back-propagation
  5. METHODOLOGY
  6. Problem Definition of DAMap
  7. Overview of Elastic Memory Network
  8. Memory Encoding via Random Projection
  9. Distributed Memory Storage
  10. Confidence based Memory Retrieval
  11. Reinforced Memorization for Domain Adaptation
  12. Experiments
  13. Experimental Setup
  14. Experimental Results
  15. Parameter Analysis
  16. ...and 5 more sections

Figures (15)

  • Figure 1: The framework of EMN including the memory storage and the memory retrieval stages compared with brains.
  • Figure 2: Accuracy versus the average domain adaptation time per instance in different DAMap methods on the $S \rightarrow R$ task of the VisDA-C dataset.
  • Figure 3: The accuracy in different epochs of domain adaptation on the $P\rightarrow A$ task of the Office-Home dataset.
  • Figure 4: The updating of the memory unit on a neuron while performing the domain adaptation from MNIST to USPS in the Digits dataset.
  • Figure 5: Accuracy versus the number of Hub/Bridging nodes in the Office-31 dataset
  • ...and 10 more figures