Reproduce “Forecasting Whole-Brain Neuronal Activity from Volumetric Video” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Alexander Immer, Jan-Matthis Lueckmann, Alex Bo-Yuan Chen, Peter H. Li, Mariela D. Petkova, Nirmala A. Iyer, Aparna Dev, Gudrun Ihrke and 9 more
The authors forecast larval zebrafish brain activity directly from three-dimensional calcium-imaging videos, preserving spatial relationships discarded by trace-based models. Their adapted U-Net predicts future frames conditioned on lead time, then applies neuron masks to evaluate activity traces. With four input frames, it lowers one-step-ahead MAE by about 8 percentage points versus TSMixer on the test set, but not in every condition.
Key insights
- The performance gains of the volumetric video model primarily stem from its ability to utilize multivariate, cross-cell information and the spatial relationships between segmented neurons, rather than detailed intracellular signal distributions or unsegmented voxels.
- There is an observed trade-off between spatial and temporal context, where larger spatial contexts are crucial for short-term predictions, but extensive spatial context can lead to overfitting when ample temporal information is available.
- Surprisingly, very high input resolution (native vs. 2x or 4x downsampled) provides minimal to no benefit for forecasting accuracy, suggesting that fine-grained intracellular details in raw images may not contain useful predictive information.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.