| 配分額 *注記 |
12,220千円 (直接経費: 9,400千円、間接経費: 2,820千円)
2018年度: 6,110千円 (直接経費: 4,700千円、間接経費: 1,410千円)
2017年度: 6,110千円 (直接経費: 4,700千円、間接経費: 1,410千円)
|
|---|
| 研究実績の概要 |
The research carried over during this fiscal year has led to the refinement and optimization of computational tools for the analysis of large-scale neural recordings and to advance our understanding on fundamental coding questions on sensory-based decision making.
We have progressed on three computational frameworks for the modeling of widefield data from the mouse occipital cortex during goal-directed behaviors. Recordings consisted on GCaMP signals across 10-12 visual cortical areas as the mouse performed in a two-alternative forced choice orientation discrimination task (Abdolrahmani et al., bioRxiv 2019). The first framework is a recurrent neural network using Hessian-free (FORCE) optimization for back-to-back learning of behavioral and neural data. The RNN has been developed in Matlab (Mathworks) and aimed to closely mimic basic functional principles of cortical connectivity (including a 4:1 ratio of excitatory and inhibitory units). The second framework is an adaptation to our data of a published variational autoencoder (LFADS, Pandarinath et al., 2018), a potent machine-learning tool aiming to uncover the latent dynamics of possibly highly non-linear dynamical systems. Development was done under the TensorFlow programming environment. Finally, under the PyTorch environment we have developed an agile, vanilla RNN currently used to tackle computational questions on decision-making under uncertain input-evidence conditions. The latter framework is also being developed for correlative analysis of widefield GCaMP data as described in the previous report.
|
