3 次元ニューロモルフィックネットワークのためのランダムウェイト消去を用いたスマートアーキテクチャ
アリ エミリアーノ ホセ
2021 年度 卒 ／修士（情報科学）
修士論文の概要
Artificial Neural Networks (ANNs) are being used in many areas of applications in the information science field with the software version of them being the most widely used and, as a consequence, the most widely studied and optimized. Due to this, hardware ANNs still have a lot of room to grow and change in order to achieve a widespread use since dealing with physical devices albeit have some advantages regarding the processing speed, the cost and space limitations are the main issue when compared to their software counterparts. Different devices are being used for the hardware implementation of networks and the use of electropolymerization of conductive wires is one of the promising ones. These devices, called Molecular Synapses onward, have two advantages that make them attractive. First, no pre-fabrication is needed apart from passive electrodes, the wire is then grown between two electrodes in order to create a conductive path. This means that if the network were to use a low number of connections, no waste is generated since only the needed material would be grown. Second, the devices are not limited to 1 dimensional space, currently 2 dimensional devices have been used for the implementation of working neural networks and the possibility of working 3 dimensional neuromorphic devices exists. Although the 3D connectivity is interesting in order to mimic our brain structure, this is no easy task when an additional problem is presented: in hardware ANNs is the duplication of connections due to the impossibility of creating negative conductances, and thus the need of differential structures to represent each weight. This results in denser connections and a higher chance of wires crossing. In this thesis, 2D experiments of Molecular Synapses are shown and explained in order to set the basis of the current working devices and the problems that they have. From that basis several 3D implementation proposals will be showcased with their possible issues associated to the electrode fabrication and wire growth. Once that the 3D devices have been explained as a possibility, then a smart architecture using random weight elimination for the reduction of the number of the connections in a network will be explained for the general case and then applied to a 3D simulation case to show the possible density reduction.