IC-1201CLIP-ViT-L/14 image features support 200-way zero-shot EEG-based object recognition better than ViT-B/16 or ResNet-50 features when used as a frozen encoder in a contrastive learning framework

Yonghao Song, Bingchuan Liu, Xiang Li, Nanlin Shi, Yijun Wang, Xiaorong Gao

SourceDecoding Natural Images from EEG for Object Recognition

The paper uses three released image encoders as frozen feature extractors in a contrastive learning pipeline that aligns image features with EEG features via cosine similarity. CLIP-ViT-L/14 (pre-trained on 400M image-text pairs) achieves the highest average top-1 accuracy of 13.8% and top-5 of 39.5% on the 200-way zero-shot task. Pre-trained ViT-B/16 achieves 12.1% top-1 (1.7% lower, p > 0.05), and pre-trained ResNet-50 achieves 8.8% top-1 (6.9% lower, p < 0.001). Notably, pre-training on ImageNet-21k+1k improves ViT-B/16 by 4.2% over random initialization (p < 0.01), but pre-training on ImageNet-1k degrades ResNet-50 by 1.9% (p < 0.05). In the appendix, EVA-CLIP pre-trained on LAION-2B further improves to 18.7% top-1 and 47.6% top-5.

Evidence
correlational
Key metric
CLIP top-1 13.8%, top-5 39.5%; ViT-B/16 (pre-trained) top-1 12.1%, top-5 29.4%; ResNet-50 (pre-trained) top-1 8.8%, top-5 17.3%; ViT-B/16 (non-pre-trained) top-1 7.9%; ResNet-50 (non-pre-trained) top-1 6.9%; EVA-CLIP top-1 18.7%, top-5 47.6%
Caveat
The image encoder is frozen and only the EEG encoder is trained; the comparison isolates the image encoder's contribution but the overall system performance also depends on the authors' TSConv EEG encoder. The ViT vs CLIP gap is not statistically significant (p > 0.05).
Model
CLIP / CLIP-ViT (LC), ViT, ResNet / ResNet-152 / ResNet-101 / ResNet-50-BN, EVA-CLIP
Concepts
Linear representation
Datasets
Gifford et al. EEG Dataset [eval]
Methods
Cosine similarity / Cosine similarity analysis / Cosine semantic similarity / cosine similarity of hidden states / Sample-wise cosine similarity / Cosine similarity of attention maps / Cosine similarity perturbation analysis / Cosine similarity template matching / Cosine similarity to neighbours / Semantic consistency (cosine similarity) [eval]
Related work
BRAVL [compared-to]
Extraction
automatic-extraction