NICE

Decoding Nature Images from EEG for Object Recognition [ICLR2024]

Core idea: basic constrastive learning for image and EEG. Interesting analysis from neuroscience perspective! 🤣

p.s. We trained the base framework (NICE), that with self-attention (NICE-SA), and with graph attention (NICE-GA) five times each for the averaged results in Table 2&3.

Abstract

• Propose a self-supervised framework for EEG-based object recognition with contrastive learning, achieving remarkable zero-shot performance on large and rich datasets.
• Demonstrate the feasibility of investigating image information from EEG signals, by resolving brain activity from temporal, spatial, spectral, and semantic aspects.
• Apply two plug-and-play modules to capture spatial correlations among EEG channels, offering evidence that the model discerns the spatial dynamics of object recognition.

Datasets

many thanks for sharing good datasets!

1. Things-EEG2
2. Things-MEG (updating)

EEG pre-processing

Script path

• ./preprocessing/

Data path

• raw data: ./Data/Things-EEG2/Raw_data/
• proprocessed eeg data: ./Data/Things-EEG2/Preprocessed_data_250Hz/

Steps

1. pre-processing EEG data of each subject

   • modify preprocessing_utils.py as you need.
     • choose channels
     • epoching
     • baseline correction
     • resample to 250 Hz
     • sort by condition
     • Multivariate Noise Normalization (z-socre is also ok)
   • python preprocessing.py for each subject.

2. get the center images of each test condition (for testing, contrast with EEG features)

   • get images from original Things dataset but discard the images used in EEG test sessions.

Image features from pre-trained models

Script path

Now we release the image features extracted with CLIP model in ./dnn_feature/.

• ./dnn_feature_extraction/

Data path (follow the original dataset setting)

• raw image: ./Data/Things-EEG2/Image_set/image_set/
• preprocessed eeg data: ./Data/Things-EEG2/Preprocessed_data/
• features of each images: ./Data/Things-EEG2/DNN_feature_maps/full_feature_maps/model/pretrained-True/
• features been packaged: ./Data/Things-EEG2/DNN_feature_maps/pca_feature_maps/model/pretrained-True/
• features of condition centers: ./Data/Things-EEG2/Image_set/

Steps

1. obtain feature maps with each pre-trained model with obtain_feature_maps_xxx.py (clip, vit, resnet...)
2. package all the feature maps into one .npy file with feature_maps_xxx.py
3. obtain feature maps of center images with center_fea_xxx.py
   • save feature maps of each center image into center_all_image_xxx.npy
   • save feature maps of each condition into center_xxx.npy (used in training)

Training and testing

Script path

• ./nice_stand.py

Visualization - updating

Script path

• ./visualization/

Steps

Citation

Hope this code is helpful. I would appreciate you citing us in your paper. 😊

@inproceedings{song2024decoding,
  title = {Decoding {{Natural Images}} from {{EEG}} for {{Object Recognition}}},
  author = {Song, Yonghao and Liu, Bingchuan and Li, Xiang and Shi, Nanlin and Wang, Yijun and Gao, Xiaorong},
  booktitle = {International {{Conference}} on {{Learning Representations}}},
  year = {2024},
}