Calcium imaging provides a valuable window into neural activity, providing insight into the complex dynamics of brain function. This paper introduces a hierarchical decomposition framework to analyze calcium imaging data, using neural non-negative matrix factorization (neural NMF). The proposed method models raster plots as multilevel ensembles, decomposing high-dimensional neural activity into interpretable spatial and temporal patterns. Using calcium imaging videos recorded in one of the coauthor’s laboratories, we demonstrate how the hierarchical approach captures distinct neural motifs, uncovering spatial organization and temporal dynamics. Visualization techniques, including heatmaps, offer intuitive representations of patterns across layers, highlighting their biological relevance. This work emphasizes the potential of hierarchical decomposition in advancing our understanding of neural systems and provides a foundation for future applications, including closed-loop interventions for optogenetic studies.

Mutual information has been actively investigated as a tool for analyzing neural networks' behavior, most notably the information bottleneck theory. However, esti-mating mutual information is a notoriously tricky task, especially for high-dimensional stochastic variables. Recently, mutual information neural estimation (MINE) was proposed as a non-parametric method to estimate mutual information for contin-uous variables without discretization. Unfortunately, MINE also produces signifi-cant errors for high-dimensional variables. Analyzing the activity of neural networks requires a dimensionality reduction mechanism, with the resulting low-dimensional representations retaining as much information as possible. We investigated different dimensionality reduction methods to determine their capabilities in estimating mutual information regarding the activity of trained neural networks. We combined MIME with principal component analysis (PCA), a convolutional neural network (CNN), and global average pooling (GAP). The experiments showed that introducing dimensionality reduction provides more stable results than the baseline method. In terms of stability, PCA-MINE and GAP-MINE performed better than CNN-MINE. They also require much less computation time than CNN. Another advantage of GAP-MINE is that it requires no hyperparameter optimization. However, the results suggested that GAP-MINE may underestimate mutual information.

Neural architecture search (NAS) is a promising approach to fully automated machine learning that applies to arbitrary tasks. However, existing NAS methods tend to make a complex network that requires much time to train and predict. Since machine learning applications are now often run on mobile devices with limited computational resources, there is a need to obtain small, efficient, yet effective network models. This paper proposes MIA-NAS (Mutual Information Analysis Neural Architecture Search), an evolutionary NAS algorithm that extends Neural Architecture Search by Hill Climbing (NASH) by Elsken et al.. The proposed method introduces a procedure that removes layers that are less contributing to making predictions. The amount of contribution from each layer is quantified by estimating mutual information they have with true labels. The experiments showed that MIA-NAS produces smaller networks with comparative performance to the ones made by NASH. Besides, removing layers based on mutual information resulted in better performance than randomly selecting layers to be removed.

Health screening is conducted in numerous countries to observe general health conditions. Machine learning has been applied to health screening records to predict asymptomatic patients' future medical states. However, for medical researchers and physicians, it is crucial to know why machine learning methods made such predictions to understand the underlying mechanism of the disease and prescribe treatments; therefore, predictions must be interpretable. We investigated the ability of an attentional neural network that processes tabular data, namely TabNet, to determine attributes that contribute to making predictions of the aggravation of type 2 diabetes. We used both model-agnostic and model-specific interpretation methods. For the former, we tested SHapley Additive exPlanations (SHAP). For the latter, we used model-specific feature importance and the mask in the attentive transformer of TabNet. We found that this mask provides useful information regarding which items in a biochemical analysis affect the aggravation of type 2 diabetes. The results from model-agnostic and model-specific methods were consistent.

In many of the existing alpha matting implementations, an intermediate representation called a trimap needs to be created manually. To automate the process, we propose a generic neural network for trimap generation based on saliency map detection. Our model multi-modally learns a saliency map and a trimap simultaneously. Because of this structure, the network focuses on reducing the error of the trimap especially within the areas with high salience. We used both the saliency map detection dataset and the alpha matting dataset to achieve accuracy in extracting subjects from natural images and generating trimaps. Experiments showed that our model could generate trimaps that are almost identical to manually generated ones. The method can also be easily combined with existing alpha matting algorithms.

Deep neural networks have recently been showing great potential on visual recognition tasks. However, it is also considered difficult to tune its parameters, and it has high training cost. This work focuses on analysis of several learning methods and properties of multinomial logistic regression deep convolutional network. We implemented a scalable deep neural network, compared the efficiency of different methods and how parameters affect the learning process. We propose an efficient method of performing back-propagation with limited kernel functions on GPU and achieved better efficiency. Our conclusions can be applied to train deep networks more efficiently. We achieved the recognition rate of over 0.95 without image preprocessing and fine tuning, within 10 minutes on a single machine.

Estimating the connectivity strength based on the signals observed at each node of a network is an important task in neural network analysis. One notable example is estimating the connectivity of a biological neural network using spikes (i.e., action potentials) observed at electrodes. The research presented in this paper introduces a novel method that estimates the underlying connectivity of a given neural network based on a similarity measure applied to spike trains. Specifically, we use a normalized positive definite kernel defined on spike trains to estimate network connectivity. The proposed method was evaluated in the context of synthetic and real data. The generation of synthetic data is based on a CERM (Coupled Escape-Rate Model), which is known to generate spike trains of various types by tuning a few parameters. We also analyzed real data recorded from the visual cortex of an anaesthetized cat. The results showed that our method provides an effective way of estimating connectivity when spike trains are the only observable information.

Concentration is one of the most important factor in determining the efficiency of learning. There has not been, however, much systematic research on controlling the level of concentration. We therefore examined the effect of an external factor, namely playing music, on the performance on a task that requires much attention. We compared three conditions: music that the subject likes, music that the subject is not familiar with, and silence. The result showed that listening to music that the subject likes do increase the performance level. Also, we discovered that there exist different temporal patterns in the change of performance. The result also indicated relationship between the temporal pattern in concentration and the external factor.

There is a growing interest in analyzing multineuron spike trains, which are spike timing data obtained from multiple neurons in the brain. Kernel methods have been successful in clustering and classification of single-neuron spike trains. We extend these methods to multineuron spike trains. Among various possible extensions, the mixture kernel was found to be most effective. The optimum parameter obtained from training this kernel was close to a biologically plausible value, suggesting that our approach is effective for seeking an appropriate model for the activity of a set of neurons.

With the growing amount of information on the WWW and the improvement of mobile computing environments, mobile Web search engines will increase significance or more in the future. Because mobile devices have the restriction of output performance and we have little time for browsing information slowly while moving or doing some activities in the real world, it is necessary to refine the retrieval results in mobile computing environments better than in fixed ones. However, since a mobile user’s query is often shorter and more ambiguous than a fixed user’s query which is not enough to guess his/her information demand accurately, too many results might be retrieved by commonly used Web search engines. This paper proposes two novel methods for query modification based on real-world contexts of a mobile user, such as his/her geographic location and the objects surrounding him/her, aiming to enhance location-awareness, and moreover, context-awareness, to the existing location-free information retrieval systems.

We define "Secure Space" as a physical space in which any resource is always protected from its unauthorized users in terms of enforcing its authorization policies assuredly. Aiming to build such secure spaces, this paper proposes an architecture and a model for space entry control based on its dynamically changing contents, such as users, physical resources and virtual resources outputted by some embedded devices. We first describe the architecture and then formalize the model and mechanism for secure spaces.

Integration of Web search with geographic information has recently attracted much attention. There are a number of local Web search systems enabling users to find location-specific Web content. In this paper, however, we point out that this integration is still at a superficial level. Most local Web search systems today only link local Web content to a map interface. They are extensions of a conventional stand-alone geographic information system (GIS), applied to a Web-based client-server architecture. In this paper, we discuss the directions available for tighter integration of Web search with a GIS, in terms of extraction, knowledge discovery, and presentation. We also describe implementations to support our argument that the integration must go beyond the simple map-and hyperlink architecture.