Closed ZiyiTsang closed 1 month ago
Thanks for your interests in our work.
Please refer to the last link of the Broader Information section in README.md.
Since the output of the transformer blocks for masked units is used as supervision for latent masked reconstruction, we hope that its output can stably represent signal features and should not fluctuate too much by just several updating steps during training. Therefore, the moving average is crucial as it allows the transformer blocks for masked units to be more influenced by historical gradients, leading to stable but continuously improved EEG feature representations over long pre-training periods.
That make sense. Thanks for your answering and looking forward to your future works!
I would like to ask one more question, when I am training the freEncoder, the test accuracy can only reach around 0.2-0.3, though it is much better than random classifier(0.05 for 40 class classification), but still very low. Is it normal situation? Will it affect the image generation later?
Yes. The frequency branch is just a complement to the features. The next stage is joint fine-tuning with the TimeEncoder, which produces the final EEG embedding. The accuracy will be different I think.
BTW, we will upload a new version of the paper as soon as possible to correct some issues in the current version on arXiv and optimize the experimental content. The issues you encountered were not highlighted as a focus of ablation analysis in the arXiv version, and we apologize for any misunderstanding this may have caused.
Thanks for your kindly answering!
Hi there, to begin with, your paper and code is really impresive to me. I have fully read your paper and now I am running on your code given. However, I have some comfusion to ask regarding to the self-training of timeEncoder in your code.
Hope you can clarify and thanks for your effort to our community.