Meta-Learning for Decoding Neural Activity Data With Noisy Labels.

Area of Science:

• Neuroscience
• Machine Learning
• Computational Biology

Background:

• Neural decoding often suffers from label noise due to manual annotation, hindering model performance.
• Improving the generalizability and robustness of neural decoding models is crucial for reliable behavioral variable prediction.

Purpose of the Study:

• To develop and evaluate a deep neural network-based sample reweighting method to address label noise in neural decoding.
• To enhance the accuracy and robustness of neural decoding models in the presence of mislabeled training data.

Main Methods:

• A deep neural network approach was employed for sample reweighting.
• A small, clean validation dataset was utilized to guide the learning process and reweight training samples.
• The method was tested on simulated neural activity and calcium imaging data from the anterior lateral motor cortex.

Main Results:

• The proposed method accurately predicted behavioral variables from simulated data with up to 36% mislabeled samples.
• For anterior lateral motor cortex data, the method achieved an F1 score of approximately 0.85 even with 48% mislabeled training samples.
• Demonstrated significant improvement in model performance despite substantial label noise.

Conclusions:

• The sample reweighting method effectively mitigates the negative impact of label noise in neural decoding.
• This approach offers a robust solution for improving neural decoding model generalizability and accuracy in real-world noisy datasets.