TAIL: Task-specific Adapters for Imitation Learning with Large Pretrained Models

Authors: Zuxin Liu, Jesse Zhang, Kavosh Asadi, Yao Liu, Ding Zhao, Shoham Sabach, Rasool Fakoor

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our extensive experiments comparing prevalent parameter-efficient fine-tuning techniques and adaptation baselines suggest that TAIL with Lo RA can achieve the best post-adaptation performance with only 1% of the trainable parameters of full fine-tuning while avoiding catastrophic forgetting and preserving adaptation plasticity in continual learning settings.
Researcher Affiliation Collaboration 1Carnegie Mellon University, 2 University of Southern California, 3 Amazon Web Services
Pseudocode No The paper includes architectural diagrams and mathematical equations, but no explicitly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not contain any explicit statement or link indicating that the authors are releasing their source code for the methodology described in the paper.
Open Datasets Yes We utilize the LIBERO robotic manipulation continual learning benchmark (Liu et al., 2023a), which features a diverse range of tasks...
Dataset Splits Yes Each task provides 50 successful human demonstrations. These are divided into 40 training trajectories and 10 for validation.
Hardware Specification Yes Our experimental platform was powered by an AMD EPYC 7R32 CPU running Ubuntu 20.04.06. All trainings utilized 8 NVIDIA A10G GPUs, each with a memory of 22731 Mi B, equipped with driver version 470.199.02 and CUDA version 11.4.
Software Dependencies Yes Our experimental platform was powered by an AMD EPYC 7R32 CPU running Ubuntu 20.04.06. All trainings utilized 8 NVIDIA A10G GPUs, each with a memory of 22731 Mi B, equipped with driver version 470.199.02 and CUDA version 11.4.
Experiment Setup Yes The environment configuration and the temporal decoder (GPT-2) hyperparameters are presented in Table 4. The detailed hyperparameters are presented in Appendix B.