## Baselines To lower the entry barrier, we provide these baselines with complete code, dataloaders, and documentation: 1. **Linear stimulus reconstruction (SR)**: Learns a spatiotemporal filter to reconstruct the attended envelope from EEG; compares with candidate envelopes via correlation. A strong, interpretable baseline. This decoder is also known as: Least Square (LS), Wiener Filter (WF), Multichannel Wiener Filter (MWF), backward temporal response function (TRF). 2. **Linear SR + causal HMM**: Integrates per-window SR correlations over time via a causal HMM (Hidden Markov Model), boosting accuracy with a modest latency trade-off. HMM postprocessing is taken from: [here](https://doi.org/10.48550/arXiv.2506.24024). Please keep in mind that the HMM postprocessing has causal and non-causal versions. Please obey the **1-s look ahead restriction**. 3. **VLAAI DNN**: A stacked-convolution-block DNN that predicts the envelope from EEG. A simple, reproducible deep baseline. **Attention**: For participants who plan to train a spatial sAAD decoder: Once you have downloaded the training datasets (which will be available once Development Phase begins), please keep in mind to convert the attended azimuth to a 0-based binary label system, with `0` representing attended to left, and `1` to the right. The attended azimuth is negative if attended to left. The baseline example code is not a spatial decoder, and therefore no conversion function is provided. ## Development environments To use the linear SR, optionally with causal HMM codes, you should have access to MATLAB. Follow the instructions in the baseline to install pre-requisite. We do not provide the linear SR (and with HMM) in Python or any other programming language. Take your own responsibility to transcibe them. To use the DNN baseline codes, you should have access to a Python 3.12 or higher, and also GPU acceleration card(s) may boost your training and inference. The DNN baseline depends on [SuperHuge](https://github.com/SeanZhang99/SuperHugeAAD), a python package for cross-dataset EEG-AAD training and inference. This package is currently not available in pip, and you can access this package via github. The package requires PyTorch and PyTorch Lightning to run. In case you can use a docker image, we have prepared a Ubuntu 22.04 image with the DNN baseline and its dependency installed. You can access the docker image from [here](https://hub.docker.com/repository/docker/yuanmingzhang/icassp2027eegsaadgc/general). Please make sure to pull the latest docker image, and pull the latest Github repo. Changes may be made after the first release of the docker image. For those who do not have access to a licensed MATLAB and do not wish/unable to transcribe the MATLAB baseline into other programming language you can use, follow DNN baseline's dependency installation instruction, and use the basic linear decoder model in the SuperHuge package. Since SuperHuge is aimed for DNN training, it does not perfectly adopt a linear decoder, and training/inference speed and accuracy may differ from the MATLAB baselines. **Use it in your own responsibility.**