[논문 리뷰] Causal Induction from Visual Observations for Goal Directed Tasks

Suraj Nair, Yuke Zhu외 2명|arXiv (Cornell University)|2019. 10. 03.

Multimodal Machine Learning Applications참고 문헌 45인용 수 45

한 줄 요약

The paper presents iterative causal induction from visual observations and an attention-based goal-conditioned policy to enable agents to complete multi-step, goal-directed tasks in environments with unseen causal structures.

ABSTRACT

Causal reasoning has been an indispensable capability for humans and other intelligent animals to interact with the physical world. In this work, we propose to endow an artificial agent with the capability of causal reasoning for completing goal-directed tasks. We develop learning-based approaches to inducing causal knowledge in the form of directed acyclic graphs, which can be used to contextualize a learned goal-conditional policy to perform tasks in novel environments with latent causal structures. We leverage attention mechanisms in our causal induction model and goal-conditional policy, enabling us to incrementally generate the causal graph from the agent's visual observations and to selectively use the induced graph for determining actions. Our experiments show that our method effectively generalizes towards completing new tasks in novel environments with previously unseen causal structures.

연구 동기 및 목표

• Motivate enabling agents to perform goal-directed tasks by reasoning about latent causal structures.
• Propose a two-stage meta-learning framework: causal induction to build a DAG of macro-variables from observations, and causal inference to guide a goal-conditioned policy.
• Develop an iterative causal induction network with attention to incrementally update causal graphs from interactive data.
• Introduce an attention-based graph encoding in the policy to focus on relevant causal edges at each step.
• Show that factorizing induction and inference via causal graphs generalizes to unseen structures with limited training cases.

제안 방법

• Iterative causal induction network F ￰fff to construct a DAG \u001b[Chat{C}] from a trajectory of visual observations and actions.
• Edge Decoder outputs edge updates \u0003bDelta e and an attention vector to apply updates to graph nodes.
• Attention bottleneck in the policy \u001b[alpha] focuses on edges relevant to the current step for action selection.
• Policy ￿gamma_G(s,g,¨C) uses attention over the causal graph to select edges and produce actions.
• Training uses supervised learning to minimize L2 loss between ground-truth and predicted C, and DAgger to train the policy with oracle guidance.

실험 결과

연구 질문

• RQ1 Can an iterative, attention-guided causal induction network accurately recover underlying causal graphs from visual interaction data?
• RQ2 Does an attention bottleneck in the goal-conditioned policy improve generalization to unseen causal structures?
• RQ3 Can combining iterative causal induction with an attention-based policy outperform prior work on visual goal-directed tasks with novel causal relations?
• RQ4 How does performance vary with the number of seen training causal structures and task size?

주요 결과

• The iterative induction network with attention (ICIN) outperforms non-iterative and ablated variants in recovering causal graphs (F1 score on unseen structures).
• The policy with an attention bottleneck (ICIN) achieves higher success rates on unseen causal structures across various switch counts and structure types compared to baselines.
• ICIN nearly matches Oracle performance in the 5-switch, 50-seen-structures setting, indicating strong causal graph induction.
• Attention bottlenecks in the policy yield significantly better generalization, with roughly 10 percentage-point gains in 1:1 and Masterswitch cases and about 40 points in 1:K and K:1 cases.