|
内容記述 |
Introduction: Detecting covarying signal structures in multi-dimensional time series data is a central task in network science. In neuroscience, for example, the identification of a group of neurons exhibiting simultaneous (de)activation provides an important clue to understanding the organization of a functional brain. Linear dimension reduction techniques, such as principal component analysis (PCA), independent component analysis (ICA), and non-negative matrix factorization, are typically utilized to estimate latent co-activation patterns embedded in a given dataset. These standard methods, however, assume that the latent components take continuous values, which may be suboptimal for representing discrete dynamics that involve system-wide, multiple co-activation mechanisms. Here, we have established an Ising machine-based analytical framework that optimizes the parameters of a linear fitting model, including binary latent components, which better represent the co-activation nature in a given time series. This method, termed binary component analysis (BiCA), can be regarded as a variant of PCA with binary latent components. Unlike standard linear dimension reduction methods with continuous latent components, gradient descent methods cannot be applied to the optimization of BiCA because the components to be estimated are discrete. In our proposed algorithm, the estimation of latent components is achieved by utilizing Ising machines and quantum annealers, which have recently garnered significant attention for their computational utility in solving combinatorial optimization problems.Results: Using simulation data, we confirmed that the proposed BiCA method reliably estimated co-activation patterns with accuracy higher than the existing methods described above. When applied to a neural activity dataset recorded in a mouse brain using two-photon mesoscopy, BiCA successfully identified the network of neurons that selectively activated in response to air-puff whisker stimulation. These results suggest that our estimation method provides beneficial insights for understanding and interpreting information processing in a target system such as the brain.Acknowledgements: This work was supported by MEXT Q-LEAP (JPMXS0120330644), JSPS KAKENHI (JP24K02341), JST ERATO (JPMJER1801), and JST PRESTO (JPMJPR2128). |
