You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
During the last ICLR 2024, I had some interesting discussions, and some people close to application with EEG in real life/VR asked for more tutorials close to real application.
I thought it was a fair request. Maybe engage more company's people to contribute.
Maybe something with continuous learning or using some probability output to adjust the model in a more pseudo-online approach.
I don't know what is the best approach to go deeper into this question. Do you have any ideas or suggestions for @agramfort, or @hubertjb (tagging to get ideas or suggestions)?
The text was updated successfully, but these errors were encountered:
I think an okay way to go could be optimizing the speed (model perspective) to do something close to a real application.
Highlight the models' size, explain how to quantize the input with different precisions, and use pred_prob. A good model in real life needs to run on very limited hardware, and I think it is a common bottleneck.
But I am not currently working with this, so I'm speculating what might be more interesting.
So with real-life you mean like something for a complete BCI system? I remember from our BCI one thing we did was finetuning online, so given an already pretrained model, given some new data, first predict it (to control sth), but then when a trial is completed use the data (and all other trials up to that point) to finetune for the new sesssion/subject... so something like this might be possible
During the last ICLR 2024, I had some interesting discussions, and some people close to application with EEG in real life/VR asked for more tutorials close to real application.
I thought it was a fair request. Maybe engage more company's people to contribute.
Maybe something with continuous learning or using some probability output to adjust the model in a more pseudo-online approach.
I don't know what is the best approach to go deeper into this question. Do you have any ideas or suggestions for @agramfort, or @hubertjb (tagging to get ideas or suggestions)?
The text was updated successfully, but these errors were encountered: