Mixing Natural and Synthetic Images for Robust Self-Supervised Representations

This paper introduces DiffMix, a new self-supervised learning (SSL) pre-training framework that combines real and synthetic images. Unlike traditional SSL methods that predominantly use real images, DiffMix uses a variant of Stable Diffusion to replace an augmented instance of a real image, facilitating the learning of cross real-synthetic image representations. The key insight is that while SSL methods trained solely on synthetic images underperform compared to those trained on real images, a blended training approach using both real and synthetic images leads to more robust and adaptable representations. Experiments demonstrate that DiffMix enhances the SSL methods SimCLR, BarlowTwins, and DINO, across various robustness datasets and domain transfer tasks. DiffMix boosts SimCLR's accuracy on ImageNet-1K by 4.56\%. These results challenge the notion that high-quality real images are crucial for SSL pre-training by showing that lower quality synthetic images can also produce strong representations. DiffMix also reduces the need for image augmentations in SSL, offering new optimization strategies.

Parameter Efficient Fine-tuning of Self-supervised ViTs without Catastrophic Forgetting

Artificial neural networks often suffer from catastrophic forgetting, where learning new concepts leads to a complete loss of previously acquired knowledge. We observe that this issue is particularly magnified in vision transformers (ViTs), where post-pre-training and fine-tuning on new tasks can significantly degrade the model's original general abilities. For instance, a DINO ViT-Base/16 pre-trained on ImageNet-1k loses over 70% accuracy on ImageNet-1k after just 10 iterations of fine-tuning on CIFAR-100. Overcoming this stability-plasticity dilemma is crucial for enabling ViTs to continuously learn and adapt to new domains while preserving their initial knowledge. In this work, we study two new parameter-efficient fine-tuning strategies: (1)~Block Expansion, and (2) Low-rank adaptation (LoRA). Our experiments reveal that using either Block Expansion or LoRA on self-supervised pre-trained ViTs surpass fully fine-tuned ViTs in new domains while offering significantly greater parameter efficiency. Notably, we find that Block Expansion experiences only a minimal performance drop in the pre-training domain, thereby effectively mitigating catastrophic forgetting in pre-trained ViTs.