Contrastive EEG-to-CLIP Alignment on THINGS-EEG: An Ablation of Target Modality, Temperature, and Channel Count

Abstract

This deposit contains code and artefacts for an ablation study ofEEG-to-CLIP contrastive alignment on the THINGS-EEG benchmark. We train anattention-based temporal-spatial encoder to map 17-channel EEG epochs intothe CLIP embedding space and evaluate nearest-concept retrieval on 200held-out concepts. Image CLIP targets outperform text CLIP targets(4.5% vs. 3.5% Top-5); mixing the two modalities degrades performance beloweither alone; and expanding from 17 to 63 channels reduces accuracy at fixedtraining scale, implicating data scale as the dominant bottleneck. Channelattention concentrates on left parietal P3, consistent with the P300 ERPcomponent. The best configuration attains 4.5% Top-5 accuracy (1.8× chance);a one-sided binomial test yields p = 0.066 with a 95% Wilson interval of[0.024, 0.084] that includes the chance rate, so we do not claimstatistically significant alignment on this test set. The work performsretrieval, not reconstruction: EEG conditions the retrieval step; imagesynthesis is conditioned on the retrieved text label. No sleep-EEG decodingis attempted.