| 研究課題/領域番号 |
16F16734
|
|---|
| 研究機関 | 国立研究開発法人理化学研究所 |
|---|
研究代表者 |
ベヌッチ アンドレア 国立研究開発法人理化学研究所, 脳科学総合研究センター, Team Leader (50722352)
|
|---|
| 研究分担者 |
FRANDI EMANUELE 国立研究開発法人理化学研究所, 脳科学総合研究センター, 外国人特別研究員
|
|---|
| 研究期間 (年度) |
2016-07-27 – 2019-03-31
|
| キーワード | recurrent neural network / large neuronal networks / machine learning / theoretical neuroscience / visual cortex |
|---|
| 研究実績の概要 |
Our research activity mainly focused on the characterization of the dynamical properties emerging from networks of recurrently connected cells in response to external inputs. In collaboration with experimentalists in the lab, we collected and analyzed data recorded from large neuronal networks in the visual cortex of mice presented with visual stimuli (specifically, gratings of different orientations). We observed changes in stimulus selectivity which are possibly consistent with functional-specific synaptic plasticity in the cortex of the adult animals, and developed mathematical models to explain the dynamical mechanisms underlying such effects at the neural network level. An important role in this analysis was played by modern machine learning tools such as dimensionality reduction methods and classification models. Specifically, we successfully exploited a combination of demixed principal component analysis (dPCA, Brendel et al., 2010) and support vector machines (SVMs, Cortes and Vapnik, 1995) to achieve the following goals: (1) represent the evolution of neuronal responses to different inputs in an appropriate state space, to elucidate and characterize the structure of their dynamical trajectories and their dependence on different task variables (time and stimulus orientation); (2) understand how downstream brain networks use the information encoded in the neuronal activity to decode the orientation of the visual stimulus, and which components of the neural population function as computational anchors for such information.
|
| 現在までの達成度 (区分) |
現在までの達成度 (区分)
2: おおむね順調に進展している
理由
The overall study and implementation of the machine learning framework which will serve as the basis for the analysis of neural network dynamics is proceeding as planned. The implementation of a general framework based on recurrent neural network models, which constitutes the key part of the project, is currently in progress and expected to be completed in the current FY. Although our initial experiments were carried out on data of a slightly different nature from that outlined in the original proposal, the machine learning methodologies we are developing are very general and flexible, and can be readily applicable to neural recordings from different contexts that researchers at the lab are working on, including behavioral tasks involving rewards (see section 12 below). From the point of view of scientific results, the project is progressing smoothly. The results obtained during the first months of research have been presented at a flagship neuroscience conference (SfN 2016, San Diego, USA), and the results of our most recent analyses are being collected into a journal paper whose preparation is already at an advanced stage.
|
| 今後の研究の推進方策 |
Building on the results on neural information encoding in the visual cortex already obtained during the last FY, we will apply recurrent neural network models (RNNs) to analyze the computational mechanisms at play in the neuronal population, in line with our original proposal and with recent suggestions from the literature (e.g. Song et al., 2016, Chaisangmongkon et al., 2017). In accordance with the main goal of the project, we will then apply the same methodologies to large-scale neural data recorded from mice engaged in reward-driven behavioral tasks, which is already being collected and analyzed in the lab. Importantly, the obtained models will be causally validated by testing the experimental reproducibility of their predictions. This goal will be accomplished using cutting-edge optogenetic tools to selectively manipulate the network dynamics. Drawing inspiration from recent work (Miconi, 2016) and from our own studies on synaptic plasticity, we will also proceed in parallel towards the goal of enhancing our recurrent network models by means of biologically inspired learning rules.
|
