[ICPR22] Moment-based Adversarial Training for Embodied Language Comprehension

Transcript

1. Moment-based Adversarial Training for Embodied Language Comprehension Shintaro Ishikawa and Komei Sugiura Keio University
2. Target: Building robots that translate from language to actions 2 [Toyota Research Institute/YouTube] Increasing need for daily care and support in aging society Motivation Robots are expected to execute household tasks following users’ linguistic instructions e.g.) “Bring me a bottle of water in the fridge” Estimating unspecified subgoals in instructions is helpful to accomplish hierarchical tasks
3. Proposed method: HLSM-MAT (1/2) 3 HLSM-MAT: HLSM with Moment-based Adversarial Training (MAT) - Add adversarial perturbations to instruction 𝑙, past subgoals 𝑮!:#$! , and state representation 𝑉% - Update perturbations following MAT during training High-level controller HLSM: [Blukis+, CoRL’21]
4. Proposed method: HLSM-MAT (2/2) 4 1. Compute gradient of cross-entropy loss 2. Introduce two types of moving-averages 𝐸 𝜹 = CE 𝑓 𝒙 , 𝒚 ∇𝜹 𝐸 𝜹 = 𝜕𝐸 𝜕𝜹 𝒎% = 𝜌! 𝒎%$! + 1 − 𝜌! ∇𝜹 𝐸 𝜹% 𝒗% = 𝜌' 𝒗%$! + 1 − 𝜌' ∇𝜹 𝐸 𝜹% ' 3. Obtain perturbation update ∆𝜹𝒕 4. Update perturbation using ∆𝜹𝒕 6 𝒎% = 𝒎% 1 − 𝜌! % , 6 𝒗% = 𝒗% 1 − 𝜌' % ∆𝜹𝒕 = 𝜂 6 𝒎% 6 𝒗% + 𝜖 𝜹%)! = Π 𝜹 *+ 𝜹% + ∆𝜹𝒕 ∆𝜹𝒕 , Algorithm of MAT: Using moments to update perturbation 𝜹, unlike VILLA [Gan+, NeurIPS’20]
5. Experimental result on ALFRED: Our method outperformed baselines 5 Method Validation Test Unseen Seen Unseen Seen SR GC SR GC SR GC SR GC HiTUT [Zhang+, ACL’21] 10.23 20.71 18.41 25.27 11.12 17.89 13.63 21.11 LAV [Nottinghum+, 21] - - 12.7 23.4 6.3 17.3 13.4 23.2 HLSM [Blukis+, CoRL’21] 18.28 31.24 29.63 38.74 20.27 30.31 29.94 41.21 Ours 18.39 31.32 30.00 41.39 21.39 32.14 31.83 43.88 ü HLSM-MAT outperformed existing methods for all metrics on ALFRED benchmark [Shridher+, CVPR’20] ü Agent predicted appropriate subgoals in correct order and achieve tasks accurately “Place the two pillows on the sofa” “Examine a cup under a lamp”