20201022
FTR: 今泉
Research on Mold Geometry Optimization Problems Using PCA and GA
In the tire manufacturing process, it is difficult to predict the shape of the tire from the shape of the mold, so the mold and materials must be adjusted.
If predictable, it can improve production efficiency and reduce costs.
In my research, predictions are made using PCA and GA based on the rate of expansion from the mold to the tire.
In this seminar, I’ll explain the progress.
Object Detection Framework using Active Learning
Recently, the demand fo ICSI, which is one of the ART, is increasing.
In ICSI, an expert embryologist finds appropriate sperms for injection, which takes much time and effort.
This research aims at detection of sperms from the microscopic movie for efficient sperm-selection.
Today, I`m going to show new method for sperm detection using Active Learning and Fuzzy Measures.
How Powerful are Graph Neural Networks?
Graph Neural Network(GNN)がグラフの表現(埋め込み)を学習するために有効であると知られており，グラフの近傍ノードの集約(Aggregation)によりグラフノードの表現ベクトルを獲得する．
しかし，これらの手法はヒューリスティクスに求められた手法であり各々のタスクでSOTAを主張するが精度が頭打ちになっている．
そこで本論文ではGNNと類似したWLtestと呼ばれるグラフ同型判定アルゴリズムを用いてGNNの理論的分析を行い，GNNの集約に関して解明をする．
また，分析結果からGraph Isomorphism Networkを提案し，グラフ分類のベンチマークにより，従来のGNNよりも有効であることを示す．
