礒川 悌次郎

In this paper, a scheme of recognizing hematopoietic tumor patients is presented, using self-organizing maps constructed by fast block-matching-based learning. This fast learning is referred to as T-BMSOM leaning. To classify the patients, screening data of examinees are presented to a constructed map. In T-BMSOM learning, a set of neurons arranged in square is regarded as a block, and one of the blocks is chosen as a winner per the presented data. It is assumed that members of a training data set to construct the map never change in static environments, whereas the data set is suddenly updated during learning in dynamic environments. While adopting the concept of blocks makes it possible to construct well-organized maps in dynamic environments, it lengthens the time for learning. To overcome this issue, T-BMSOM learning is based on a decision-tree-like winner search and a batch process. The screening data of an examinee frequently lacks several of the item values, and hence the data is presented to the map after averages of non-missing item values substitute for items with no values. The class of the data to be classified is basically judged by observing the label of a winner block. Simulation results establish that the proposed scheme achieves high accuracy of correctly recognizing the data of hematopoietic tumor patients, even if training the map is conducted in a dynamic environment.

実験データに基づく数理モデルによる神経回路のシミュレーション解析は,脳をシステムとして解析していくための有効なアプローチの一つと考えられる.このようなアプローチにおいて,昆虫脳は比較的簡単な構造でありながら,様々な感覚受容と感覚情報に基づく行動制御機能を備えており,脳における神経回路の構造及びその情報処理メカニズムの本質を解明するためのモデル動物として極めて有用である.本研究では,カイコガ触角葉の神経細胞を対象に,共焦点レーザ顕微鏡画像から神経細胞の三次元形態モデルを再構成する手法を提案し,更に,そのモデルを用いて電気的応答特性の解析を可能とした.これによって細胞形態の正確なトレースが可能になり,神経細胞間の結合状態をより詳細に同定できるようになった.更に,神経細胞の電気的特性に関するシミュレーション解析を実施することで,細胞の形態情報に基づく神経回路の構造に加えて,応答のダィナミックスを考慮した解析を可能とした.

魚や鳥が形成する群行動は，群を構成する各エージェント（個体）に現実の生物の行動パターンを模した行動ルールを規定することでモデル化できることが知られている．本研究では，強化学習を用いた学習によりエージェント間の相互作用から群行動を自律分散的に形成させる枠組みを提案してきた．本稿では，この枠組みがさまざまな環境に適応していくモデル設計となっているか，計算機実験により群行動の発現について検討する．

Self-organizing map-based methods for the detection of confusion between blood test data are presented. Learning data for the self-organizing map (SOM) is generated by subtracting each element of complete blood count (CBC) data of the immediately previous patient's results from that of the current results. The neurons in the well-trained SOM are roughly divided into two clusters: one with neurons reacting to regular input data, and the other with neurons reacting to irregular input data generated by subtraction between confused CBC data. If a winner neuron belongs to the latter cluster, it is presumed that confusion has arisen between the CBC data of different patients. In addition, a genetic algorithm is adopted to eliminate redundant elements in the CBC data, which have an unfavorable influence on the judgment of confusion. Experimental results show that the proposed methods achieve high accuracy of detection even when the input data irrelevant to the learning of maps is applied to them.