Artificial Intelligence is not just the grand plan of the automation of statistical mod- eling, but also a daring, complicated and extremely important philosophical project that concerns the nature of our own minds and their relationship to the physical universe and the universe of meaning. What can AI contribute to our understanding of epistemology, metaphysics and philos- ophy of mind? If we ascribe a degree of intelligence to any system that can make models, what makes a system generally intelligent? Are humans in the class of generally intelligent systems themselves? Given that many functions are not learnable, does general intelligence even exist?

本稿ではまず経験から知識を獲得する知的エージェントのアーキテクチャの構想を述べる。次にそのアーキテクチャの重要な要素技術の１つとして、２層ベイジアンネットを用いてパターン獲得・パターンマッチを行う手法を提案する。本手法は特別に設計された条件付確率モデルを用いる。

Intelligent agents can significantly improve their predictive and problem-solving performanceby possessing metacognitive skills to monitor and control their own thinking, knowledge, cognition, andlearning. One of the most important abilities is introspection, which involves recall about past episodes andimagination on counterfactual time developments. Because such abilities are not yet realized by currentartificial intelligence, it is significant to refer to the brain mechanism. In recent years, there has been agrowing body of knowledge on metacognition and introspection in the brain, but the full extent of itsinformation processing is still unclear. In this paper, a working hypothesis of metacognition in the brain isproposed, based on the knowledge that representations that become conscious in the brain are independentof the working memory. The proposed model encodes a series of indices for object-level representations onthe neocortex to representations on the hippocampus, and then monitors and controls the representations onthe hippocampus from the meta-level through a part of the neocortex. Next, anatomical findings of Papezcircuit were collected, that circuit sit between the hippocampus and the higher-order cortical area, and thesefindings were organized into an information flow diagram. The retrosplenial cortex (RSC) - entorhinalcortex pathway is predicted as a plausible candidate for the transmitting metacognitive control signals fromthe neocortex to the hippocampus from the analysis of the flow diagram.

We evaluate the integrated conceptual information for a small system consistingof 6 to 8 nodes with a bridge. For numerical computations, we use the Python library termedPyPhi proposed by Tononi et al., who are proponents of the integrated information theory ofconsciousness. Our simulations show that the integrated information varies under the influence ofnetwork topology and frustration of loops. It is also found that in systems with a bridge, smallsubsystems consisting of less than half of the total number of nodes comparatively form a majorcomplex which is the combination of nodes that generate the largest integrated information.

Living organisms have evolved into better life forms while repeating mutations in response to changes in various living environments. On the other hand, the field of view of the human eye (clear visible area) is very narrow, except for the central field of view, it looks blurred, the extent to which changes / differences such as movement can be recognized. At first glance, it seems unreasonable to understand. In this paper, I introduce the advanced information processing power acquired in reverse that the field of view is narrow.

One of the objects of consciousness research is the attention, which has the property of inhibitingirrelevant information and emphasizing the processing of relevant information. It has been proposed thatclaustrum (CLA) determines which information to direct selective attention to in a top-down manner. Amodel in which the CLA regulates attentional selection from higher-order regions to sensory regions wasproposed. The CLA-mediated neuronal connections can be assumed to be a counter stream structure, butreciprocal connections between the CLA and cortical and higher order regions have been identified. In thispaper, we discuss the mechanism of attentional selection

Visualizing deep neural networks (DNN) provides an intuitive explanation for thelearned internal representation, while its evaluation is difficult. We believe that a DNN 's learningrepresentation should be evaluated by its consistency with concepts owned by human. In this study,we represent such a concepts as symbolic binary representations and distributions with variance,and investigated the consistency of a specific neuroscientific concept (P300) with the representationslearned from EEG data obtained in a P300 speller experiment. As a result, we found that theconsistency between the concept and the representation is related to the discrimination accuracyof the DNN.

Introducing spiking signals to neural network operations is an important and fundamentalfeature of the neuromorphic architectures, because of reducing energy-efficiency dramatically, and foremulating information operations of human cortex. Though, in terms of application accuracy, the SpikingNeural Network, SNN, has not been outperforming the Artificial Neural Networks, the reason of which isregarded as that the SNN is following ideas of existing Deep-Neural-Network topologies and their learningmethodologies. The outline of this situation is reviewed and studied for seeing into next deployments ofthe SNN.