A Retention-Centric Framework for Continual Learning with Guaranteed Model Developmental Safety

TL;DR

This paper introduces model developmental safety (MDS), a retention-centric framework that enforces data-dependent constraints to guarantee preservation of protected capabilities during iterative model development. It formulates MDS as a constrained optimization problem and solves it efficiently for CLIP-based image classification using a quadratic penalty method, moving-average gradient estimators, and a constraint-aware objective. The approach is augmented with task-dependent heads to boost the critical parameter delta, providing provable convergence guarantees to an epsilon-KKT point and practical improvements in target-task performance while preventing forgetting on protected tasks. Experiments on autonomous driving (BD100K) and Places365 demonstrate strong retention of protected capabilities and notable gains on new/rare target tasks, with ablations highlighting the value of external data, per-task heads, and scheduling choices. The framework offers a scalable, safety-aware path for continual learning in real-world, cost-sensitive settings.

Abstract

In real-world applications, learning-enabled systems often undergo iterative model development to address challenging or emerging tasks, which involve collecting new data, training a new model and validating the model. This continual model development process raises a significant issue that acquiring new or improving existing capabilities may inadvertently lose good capabilities of the old model, also known as catastrophic forgetting. While existing continual learning aims to mitigate catastrophic forgetting by trading off performance on previous tasks and new tasks to ensure good average performance, it often falls short in cost-sensitive applications, where failing to preserve essential established capabilities introduces unforeseen costs and risks and substantial expenses for re-improving these capabilities. To address this issue, we impose a requirement on learning systems to ensure that a new model strictly retains important capabilities of the old model while improving target-task performance, which we term model developmental safety. To ensure model developmental safety, we propose a retention-centric framework with data-dependent constraints, and study how to continually develop a pretrained CLIP model for acquiring new or improving existing capabilities of image classification. We propose an efficient constrained optimization algorithm with theoretical guarantees and use its insights to finetune the CLIP model with task-dependent heads for promoting the model developmental safety. Experiments on autonomous driving and scene recognition datasets validate the efficacy of our method.

Figures (7)

• Figure 1: Performance of recognizing 6 weather conditions for autonomous driving with two rounds of model development using new data. The Round 1 development targets at overcast and Round 2 aims to improve recognizing foggy. Base refers to the CLIP model finetuned on BDD100K data.
• Figure 2: Visualization of the learning trajectory. Each dot denotes a solution with lighter color being earlier iterations and darker being later iterations.
• Figure 3: Performance Comparison with Baselines. Dot lines represent the performance of the base model on the target task. Detailed numbers are presented in Table \ref{['tab:d_tunnel']}, \ref{['tab:d_foggy']}, \ref{['tab:d_overcast']}.
• Figure 4: (Left) Adaptive weight adjustments for each protected task during training (Targeting Foggy). Weights shown are averaged over every 600 iterations for visualization. (Right) Performance comparison with baseline RM when targeting Dresssing Room on Places365 Dataset, with 2k samples per constraint. Red line denotes base model's performance, green diamonds denote the target class. RM baseline shown is for weight $\alpha=10000$ and more plots for other weights are presented in Figure \ref{['fig:places365']}.
• Figure 5: (a) Task-dependent heads promote developmental safety. (b) Performance Comparison between constant $\beta$ and increasing $\beta$.

Theorems & Definitions (18)

• Definition 1: Model Developmental Safety (MDS)
• Lemma 1: Generalization Error of Rentention
• Definition 2
• Theorem 1
• Lemma 2
• proof
• proof
• Lemma 3
• proof
• Lemma 4