TL;DR
Fine-tuning foundation models (like CLIP) often triggers "concept forgetting," where the model gains expertise in a new task but loses its broad, pre-trained knowledge. The authors propose LDIFS ($\ell_2$ distance in Feature Space), a simple yet powerful regularization method that preserves a model's original internal representations, allowing it to learn new tasks while keeping its general-purpose capabilities intact.
What Problem Does This Paper Address?
Foundation models are expensive to train and serve as massive "knowledge stores" of real-world concepts. However, to get peak performance on a specific downstream task (like identifying satellite imagery or medical scans), researchers typically perform end-to-end fine-tuning.
The authors identify a major side effect: concept forgetting. As the model's weights shift to minimize loss on the new dataset, the internal features that previously recognized thousands of other concepts are "crippled." This is particularly problematic for continual learning, where we want to update a model with new information without overwriting everything it already knows. The paper seeks to quantify this forgetting and provide a way to prevent it without sacrificing performance on the new task.
Methods
The researchers propose a new regularizer called LDIFS (short for $\ell_2$ distance in Feature Space).
Standard fine-tuning usually just minimizes a classification loss (like Cross-Entropy). LDIFS adds a penalty term: it minimizes the $\ell_2$ distance between the features generated by the original pre-trained model and the features generated by the model currently being fine-tuned.
Key technical nuances include:
- Multi-layer Preservation: Instead of just looking at the final output layer, LDIFS concatenates features from multiple intermediate layers throughout the network. This ensures that the "building blocks" of knowledge distributed across the architecture are preserved.
- Linear Probe (LP) Benchmarking: To measure forgetting, the authors use the change in Linear Probe accuracy ($\Delta_{LP}$) on a suite of unrelated datasets. They argue that LP is a better measure than Zero-Shot (ZS) accuracy because fine-tuning often breaks the specific alignment between image and text encoders, even if the underlying image features are still present.
Datasets
The study is remarkably thorough, evaluating 10 different image classification datasets with high semantic variability:
- General/Natural: CIFAR-10, CIFAR-100, ImageNet.
- Specialized: Stanford Cars (fine-grained), EuroSAT (satellite), RESISC45 (remote sensing), GTSRB (traffic signs).
- Abstract/Other: DTD (textures), MNIST (digits), SVHN (house numbers).
They also test across various architectures including CLIP (ViT and ResNet variants), FLAVA, DINOv2, and Masked Auto-Encoders (MAE).
Key Insights and Ideas
- Features > Weights: Previous methods like L2SP tried to keep the weights of the model close to the original. The authors show that regularizing the feature space (the actual representations the model produces) is far more effective. Two models can have different weights but produce similar features; LDIFS focuses on the behavior rather than the raw numbers.
- The Importance of "Shallow" Features: A key finding is that preserving only the last layer's features is insufficient. As shown in their ablation studies, including features from earlier layers is crucial for preventing concept forgetting.
- Positive Forward Transfer: Surprisingly, in 8 out of 10 tasks, LDIFS didn't just prevent forgetting—it actually resulted in positive $\Delta_{LP}$. This means that while learning a new task (like satellite imagery), the model actually became slightly better at recognizing unrelated concepts (like cars or textures) compared to standard fine-tuning.
Key Results
- Minimal Forgetting: Across all 10 tasks, LDIFS consistently showed the lowest levels of concept forgetting. While standard fine-tuning methods often saw accuracy drops of 5-10% on unrelated datasets, LDIFS often maintained or slightly improved performance.
- Competitive Accuracy: LDIFS achieves this preservation without a "tax" on the downstream task; its accuracy on the target dataset remains competitive with (and sometimes exceeds) standard end-to-end fine-tuning.
- Continual Learning Champion: In sequences of tasks (e.g., learning SVHN, then CIFAR-10, then RESISC45), LDIFS outperformed established continual learning baselines like iCaRL and LwF in both accuracy and knowledge retention.
Limitations and Future Work
- Hyperparameter Sensitivity: LDIFS introduces a $\lambda$ hyperparameter to balance the classification loss vs. the feature preservation. While the authors found $\lambda=10$ worked well generally, optimal values might vary by task.
- Bias Percolation: The authors note that if a pre-trained model has inherent biases, LDIFS will actively work to preserve those biases in the fine-tuned version.
- LLM Extension: This work focused on vision and multi-modal models. A natural next step is investigating if feature-space preservation can solve similar forgetting issues in Large Language Models (LLMs).
- Feature Selection: The choice of which intermediate layers to use was heuristic (evenly spaced). Future work could explore more principled ways to identify which layers encode specific types of knowledge.
Summary updated