Abstract:
Large-scale recordings of neural activity and behavior open up the possibility for understanding neural mechanisms. Algorithms, such as Latent Variable Models (LVMs), successfully project the high-dimensional spatiotemporal neural recordings into interpretable low-dimensional representations from where behaviors could be decoded. Recent LVMs are based on deep learning methodology where a deep neural network is trained to reconstruct the same neural activity given as input and, as a result, build a low-dimensional latent representation. While successful, the reconstruction of the same signal is non-casual, where both past and future activity is taken as input, contradicting the causal, temporal interactions in neural systems. In this study, we propose to learn neural representations incorporating the temporal causality constraint with forecasting tasks. The forecasting task, generating future neural signals given past neural signals, could capture such a causal temporal relationship. Furthermore, we introduce a prior, which consists of pairwise neural unit spatial interaction as a dynamic graphical system, to improve the forecasting performance. Experiments demonstrate the ability of the proposed method to recover ground truth spatial interactions on synthetic datasets and yield accurate estimation of future neural dynamics.
Bio: Jingyuan is a fifth-year Ph.D. student at the University of Washington majoring in Electrical and Computer Engineering, advised by Eli Shlizerman, holding a particular interest in brain+AI, i.e., applying machine learning methods to neural system recordings for human body control or machine control and building a brain-inspired machine learning approach for fast and adaptable learning. She has been working on Graph Neural Networks for ECoG recordings learning neural representation, demonstrating the possibility of the method effectively learning of predicting future neural activity with a small dataset. Besides, Jingyuan worked on an active learning enhanced behavior recognition framework with a novel uncertainty measure.
Ref: https://openreview.net/pdf?id=7ntI4kcoqG
Title: Learning Neural Representations Compatible with Behavior Interpretation
Abstract:
Large-scale recordings of neural activity and behavior open up the possibility for understanding neural mechanisms. Algorithms, such as Latent Variable Models (LVMs), successfully project the high-dimensional spatiotemporal neural recordings into interpretable low-dimensional representations from where behaviors could be decoded. Recent LVMs are based on deep learning methodology where a deep neural network is trained to reconstruct the same neural activity given as input and, as a result, build a low-dimensional latent representation. While successful, the reconstruction of the same signal is non-casual, where both past and future activity is taken as input, contradicting the causal, temporal interactions in neural systems. In this study, we propose to learn neural representations incorporating the temporal causality constraint with forecasting tasks. The forecasting task, generating future neural signals given past neural signals, could capture such a causal temporal relationship. Furthermore, we introduce a prior, which consists of pairwise neural unit spatial interaction as a dynamic graphical system, to improve the forecasting performance. Experiments demonstrate the ability of the proposed method to recover ground truth spatial interactions on synthetic datasets and yield accurate estimation of future neural dynamics.
Bio: Jingyuan is a fifth-year Ph.D. student at the University of Washington majoring in Electrical and Computer Engineering, advised by Eli Shlizerman, holding a particular interest in brain+AI, i.e., applying machine learning methods to neural system recordings for human body control or machine control and building a brain-inspired machine learning approach for fast and adaptable learning. She has been working on Graph Neural Networks for ECoG recordings learning neural representation, demonstrating the possibility of the method effectively learning of predicting future neural activity with a small dataset. Besides, Jingyuan worked on an active learning enhanced behavior recognition framework with a novel uncertainty measure.
