Simon Kern
@skjerns.de
Sleep & Memory researcher @ CIMH Mannheim with @gordonfeld.bsky.social . Interested in replay and applied machine learning in the context of episodic and declarative memory.
MEG and Python enthusiast.
MEG and Python enthusiast.
Last week I was being trained as a #MNE-Python TrainEErs at #PracticalMEEG2025 - it was a lot of fun to look behind the scenes and learn how to run a good workshop :) thanks to @cuttingeeg.bsky.social for hosting and Marijn van Vliet and @nschawor.bsky.social for organizing this amazing workshop!
November 2, 2025 at 9:20 AM
Last week I was being trained as a #MNE-Python TrainEErs at #PracticalMEEG2025 - it was a lot of fun to look behind the scenes and learn how to run a good workshop :) thanks to @cuttingeeg.bsky.social for hosting and Marijn van Vliet and @nschawor.bsky.social for organizing this amazing workshop!
Not sure how exactly they measure multilingual capabilities, but 1b seems pretty bad - might explain the german results
June 25, 2025 at 10:45 AM
Not sure how exactly they measure multilingual capabilities, but 1b seems pretty bad - might explain the german results
>30ms
Interesting! That is also my experience. For some pilot data I've once created these heatmaps in which I plot sequenceness as a function of training time point. There's some interesting patterns in there but I'm not sure how they occur. Have not found a good expansion for what is happening
Interesting! That is also my experience. For some pilot data I've once created these heatmaps in which I plot sequenceness as a function of training time point. There's some interesting patterns in there but I'm not sure how they occur. Have not found a good expansion for what is happening
June 19, 2025 at 8:04 PM
>30ms
Interesting! That is also my experience. For some pilot data I've once created these heatmaps in which I plot sequenceness as a function of training time point. There's some interesting patterns in there but I'm not sure how they occur. Have not found a good expansion for what is happening
Interesting! That is also my experience. For some pilot data I've once created these heatmaps in which I plot sequenceness as a function of training time point. There's some interesting patterns in there but I'm not sure how they occur. Have not found a good expansion for what is happening
8/ Additionally, when re-using the sample simulation published in the TDLM methods paper, we found that replay was simulated with 2000 reactivations per minute! When simulating with more realistic 15 events per minute, even the sample code fails to reach significance.
June 16, 2025 at 7:43 AM
8/ Additionally, when re-using the sample simulation published in the TDLM methods paper, we found that replay was simulated with 2000 reactivations per minute! When simulating with more realistic 15 events per minute, even the sample code fails to reach significance.
7/ To our surprise, we needed >80 replay events per minute to reach the lowest bar of significance, with more than 120 events per minute for solid evidence. Cf. typical ripple rates in humans (as a proxy for estimating replay density) are ~15 per minute.
June 16, 2025 at 7:43 AM
7/ To our surprise, we needed >80 replay events per minute to reach the lowest bar of significance, with more than 120 events per minute for solid evidence. Cf. typical ripple rates in humans (as a proxy for estimating replay density) are ~15 per minute.
5/ We took the control resting state (from before the experiment) of each participant as a basis. We then inserted subtle patterns of each stimulus from the localizer with specific time intervals into the resting state.
June 16, 2025 at 7:43 AM
5/ We took the control resting state (from before the experiment) of each participant as a basis. We then inserted subtle patterns of each stimulus from the localizer with specific time intervals into the resting state.
4/ Our decoders work well (~35-40% accuracy/10% chance) so we decided to investigate. We came up with a hybrid simulation approach that simulates replay under more naturalistic conditions (i.e. endogenous oscillations). Previous studies only used completely synthetic data.
June 16, 2025 at 7:43 AM
4/ Our decoders work well (~35-40% accuracy/10% chance) so we decided to investigate. We came up with a hybrid simulation approach that simulates replay under more naturalistic conditions (i.e. endogenous oscillations). Previous studies only used completely synthetic data.
3/ We let 30 participants learn a hidden graph structure while in the MEG (criterion 80% memory performance). We expected to find task-related replay in a resting state session after learning. However no clear sequenceness was found, despite similar pipeline as previous papers.
June 16, 2025 at 7:43 AM
3/ We let 30 participants learn a hidden graph structure while in the MEG (criterion 80% memory performance). We expected to find task-related replay in a resting state session after learning. However no clear sequenceness was found, despite similar pipeline as previous papers.
Dall-e is really awesome for creating lecture illustrations without all the copyright-hassle🤩
November 20, 2023 at 10:55 AM
Dall-e is really awesome for creating lecture illustrations without all the copyright-hassle🤩