Low Number Of Data Points

[Barbod1380]

I have data where each recording lasts for approximately 400 ms. This results in a low number of data points. For example, if we divide the recording into 5 ms time bins, we obtain 80 data points. Can CEBRA work with such a small number of data points? Neural networks typically require a large amount of data for good performance. If CEBRA can handle this, does it require specific parameter tuning, or is a grid search necessary?

[stes]

80 datapoints in total are indeed a very low number. How many datapoints does the dataset have overall?

Regarding your question: the most important params to consider is the model size in this case (hidden dimensions, architecture). But even then, only 80 points will be generally too few if the goal is to train a nonlinear neural net encoder.

[Barbod1380]

In the dataset, each frame was presented to the mouse 30 times, with each presentation lasting 400ms (these 30 repetitions were not consecutive). For neural manifold generation, I segmented this 400ms window into 5ms time bins, resulting in 80 data points per presentation. I could use smaller time bins to increase the number of data points, but this would lead to a much sparser matrix.