• Recurrent Relational Networks

    WHY? Some tasks such as Sudoku require serial steps of relational inference. WHAT? Recurrent relational network operates on a graph representation of objects. Message passing method is used to pass the relational information to neighbor nodes to solve the task. The loss is minimized at every step. So? This module...


  • Modularity Matters: Learning Invariant Relational Reasoning Tasks

    WHY? Former CNN models fully activate(filly distributed features) for a single input showing poor performance on invariant relational reasoning. WHAT? The reason former CNN models are poor at invariant reasoning is interference problem that learning each pattern interfere with each other while a model try to learn many patterns. To...


  • Learning Visual Question Answering by Bootstrapping Hard Attention

    WHY? Hard attention is relatively less explored than soft attention. WHAT? This paper showed that hard attention can be competitive and efficient as soft attention by bootstraping hard attention. In constrast to soft attention, hard attention discretely choose the point to attend. The key idea is to use L2-Norm of...


  • Relational Deep Reinforcement Learning

    WHY? Relational information is important in some reinforcement learning tasks. WHAT? Relational information can be important source for high scores in reinforcement learning. To provide inductive bias for relational information, this paper appied self-attention(MHDPA) to the last layer of the convolution network that encode the state. Another multilayer perceptron() is...


  • Neural-Symbolic VQA: Disentangling Reasoning from Vision and Language Understanding

    WHY? Most VQA reasoning algorithms are not transparent in their reasoning and not robust to complex reasoning. WHAT? This paper combined two different methods: deep representation learning and symbolic program execution. Neural Symbolic VQA consists of three stages: scene parsing, question parsing and program execution. In scene parsing stage, Mask...