LEARNING APPARATUS, METHOD, PROGRAM, AND INFERENCE APPARATUS

• Main Authors: SASAYA TENTA, WATANABE TAKASHI, ONO TOSHIYUKI
• Format: Patent
• Language: English; Japanese
• Published: 13.06.2024
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; PHYSICS

To process data including unnecessary information with high accuracy.SOLUTION: A learning apparatus includes a first extraction unit, a second extraction unit, a conversion unit, a calculation unit, a first processing unit, and an update unit. The first extraction unit extracts a first feature quantity from input data by processing based on a parameter. The second extraction unit extracts a second feature quantity from the input data by processing based on a parameter. The conversion unit stochastically converts at least one of the first feature quantity and the second feature quantity to generate a first converted feature quantity and a second converted feature quantity. The calculation unit calculates a first loss regarding similarity between the first converted feature quantity and the second converted feature quantity. The first processing unit executes processing based on parameters, on the first converted feature quantity to obtain a first processing result. The update unit updates at least one parameter of the first extraction unit and the first processing unit so that a value based on the first loss and a second loss calculated from the first processing result and ground truth data is minimized.SELECTED DRAWING: Figure 1 【課題】不要な情報を含むデータを高精度に処理できる。【解決手段】本実施形態に係る学習装置は、第１抽出部と、第２抽出部と、変換部と、計算部と、第１処理部と、更新部とを含む。第１抽出部は、パラメータに基づく処理により入力データから第１特徴量を抽出する。第２抽出部は、パラメータに基づく処理により入力データから第２特徴量を抽出する。変換部は、第１特徴量と第２特徴量との少なくとも一方を確率的に変換し、第１変換特徴量および第２変換特徴量を生成する。計算部は、第１変換特徴量および第２変換特徴量との類似度に関する第１ロスを計算する。第１処理部は、第１変換特徴量に対してパラメータに基づく処理を実行し、第１処理結果を得る。更新部は、第１ロスと、第１処理結果と正解データとから計算される第２ロスとに基づく値が最小となるように、第１抽出部および第１処理部の少なくとも一方のパラメータを更新する。【選択図】図１