2018/03/20 by Ronan Riochet, Mario Ynocente Castro, Riochet, Ronan +11 · 10 citations
Computer Science · Neuroscience · #Human Pose and Action Recognition #Cognitive Science and Education Research #Explainable Artificial Intelligence (XAI)
paper · pdf · doi:10.48550/arxiv.1803.07616
In order to reach human performance on complexvisual tasks, artificial\nsystems need to incorporate a sig-nificant amount of understanding of the world\nin termsof macroscopic objects, movements, forces, etc. Inspiredby work on\nintuitive physics in infants, we propose anevaluation benchmark which diagnoses\nhow much a givensystem understands about physics by testing whether itcan tell\napart well matched videos of possible versusimpossible events constructed with\na game engine. Thetest requires systems to compute a physical plausibilityscore\nover an entire video. It is free of bias and cantest a range of basic physical\nreasoning concepts. Wethen describe two Deep Neural Networks systems aimedat\nlearning intuitive physics in an unsupervised way,using only physically\npossible videos. The systems aretrained with a future semantic mask prediction\nobjectiveand tested on the possible versus impossible discrimi-nation task. The\nanalysis of their results compared tohuman data gives novel insights in the\npotentials andlimitations of next frame prediction architectures.\n