Online Analytic Exemplar-Free Continual Learning with Large Models for Imbalanced Autonomous Driving Task (2405.17779v2)

Abstract: In autonomous driving, even a meticulously trained model can encounter failures when facing unfamiliar scenarios. One of these scenarios can be formulated as an online continual learning (OCL) problem. That is, data come in an online fashion, and models are updated according to these streaming data. Two major OCL challenges are catastrophic forgetting and data imbalance. To address these challenges, in this paper, we propose an Analytic Exemplar-Free Online Continual Learning algorithm (AEF-OCL). The AEF-OCL leverages analytic continual learning principles and employs ridge regression as a classifier for features extracted by a large backbone network. It solves the OCL problem by recursively calculating the analytical solution, ensuring an equalization between the continual learning and its joint-learning counterpart, and works without the need to save any used samples (i.e., exemplar-free). Additionally, we introduce a Pseudo-Features Generator (PFG) module that recursively estimates the mean and the variance of real features for each class. It over-samples offset pseudo-features from the same normal distribution as the real features, thereby addressing the data imbalance issue. Experimental results demonstrate that despite being an exemplar-free strategy, our method outperforms various methods on the autonomous driving SODA10M dataset. Source code is available at https://github.com/ZHUANGHP/Analytic-continual-learning.

• The paper introduces AEF-OCL, a novel exemplar-free method leveraging analytic ridge regression to prevent catastrophic forgetting in online learning.
• It incorporates a pseudo-feature generator that recursively estimates feature distributions to balance imbalanced data in real-time driving scenarios.
• Experimental results on the SODA10M dataset show an AMCA of 66.32%, outperforming traditional exemplar-based and replay approaches.

Online Analytic Exemplar-Free Continual Learning with Large Models for Imbalanced Autonomous Driving Task

The paper "Online Analytic Exemplar-Free Continual Learning with Large Models for Imbalanced Autonomous Driving Task" introduces an innovative approach to the challenges posed by online continual learning (OCL) in the context of autonomous driving. The primary focus is on addressing two significant issues: catastrophic forgetting and data imbalance, which frequently arise when models are updated with streaming data.

Methodological Advancements

1. Analytic Exemplar-Free Online Continual Learning (AEF-OCL): The authors propose AEF-OCL, which utilizes analytic continual learning principles. The approach leverages ridge regression as a classifier that works with features extracted by a large-scale pre-trained model. The recursive calculation of the ridge regression ensures the retention of previously acquired knowledge while avoiding the storage of past samples, classifying AEF-OCL as an exemplar-free method.
2. Pseudo-Features Generator (PFG): To address data imbalance, a PFG module is introduced. This module estimates the mean and variance of the features for each class recursively. It generates pseudo-features from the estimated normal distribution matching the actual features to balance the training data, thereby alleviating the problems associated with data imbalances in the training process.
3. Recursive Learning Framework: The AEF-OCL recursively updates the classifier's weights to maintain equivalence with a joint-learning scenario, ensuring no catastrophic forgetting. By using only the current task's data, the method circumvents the need for storing previous data exemplars, maintaining data privacy and reducing memory requirements.

Experimental Evaluation

The authors evaluate the performance of AEF-OCL using the SODA10M dataset, a prominent dataset for autonomous driving that exemplifies significant class imbalance. Results indicate that AEF-OCL outperforms existing methods in terms of Average Mean Class Accuracy (AMCA), achieving an AMCA of 66.32%, which surpasses both exemplar-free and some replay-based methods. This demonstrates the method's effectiveness in tackling catastrophic forgetting and data imbalance.

Theoretical Contributions

The theoretical groundwork in the paper provides proofs of recursive estimations for mean and variance, along with recursive updates for the ridge regression in an online continual learning context. These contributions underline the robustness of the approach in dynamically shifting environments common in autonomous driving scenarios.

Practical Implications and Future Directions

The AEF-OCL establishes a compelling method for deploying models in real-world autonomous driving tasks where the data stream is continuous and imbalanced. The use of a pre-trained large model ensures strong feature extraction capabilities, which are crucial for handling unforeseen or rare scenarios without human intervention. However, the method's reliance on pre-trained backbones may warrant further adaptation to specific domains or environments within autonomous driving.

Future research directions include exploring adaptive backbone networks that can dynamically adjust to new data in online settings and investigating security aspects related to adversarial robustness. Considering the privacy-preserving nature of AEF-OCL, further integration into federated learning frameworks could provide additional benefits in terms of distributed and decentralized learning.

In summary, the paper provides a comprehensive solution to critical challenges in online continual learning for autonomous driving, backed by strong theoretical insights and practical outcomes. The AEF-OCL sets a precedence for future work in the field, especially concerning exemplar-free, privacy-focused, and model-efficient learning frameworks.