Efficient Equivariant Transfer Learning from Pretrained Models
Authors: Sourya Basu, Pulkit Katdare, Prasanna Sattigeri, Vijil Chenthamarakshan, Katherine Driggs-Campbell, Payel Das, Lav R. Varshney
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 Experiments. Here, we provide experimental results for equizero and λ-equitune in 5.1 and 5.2, respectively, for all the applications described in 4. |
| Researcher Affiliation | Collaboration | Sourya Basu University of Illinois at Urbana-Champaign Pulkit Katdare University of Illinois at Urbana-Champaign Prasanna Sattigeri IBM Research Vijil Chenthamarakshan IBM Research Katherine Driggs-Campbell University of Illinois at Urbana-Champaign Payel Das IBM Research Lav R. Varshney University of Illinois at Urbana-Champaign |
| Pseudocode | No | No pseudocode or algorithm block explicitly labeled or presented in a structured algorithmic format. |
| Open Source Code | Yes | The code for this paper is available at https://github.com/basusourya/lambda_equitune. |
| Open Datasets | Yes | We first pretrain Deep Q-learning nets (DQNs) for each of the Gridworld, Cartpole, and Acrobot environments using the default architecture from Raffin et al. (2021) with 103k parameters. We evaluate the performance of our algorithm on the SCAN Dataset (Lake & Baroni, 2018). |
| Dataset Splits | Yes | We first pretrain Deep Q-learning nets (DQNs) for each of the Gridworld, Cartpole, and Acrobot environments using the default architecture from Raffin et al. (2021) with 103k parameters. We pretrained all the models using a learning rate 10 4. We used training time steps as 100k, 100k, and 70k for Gridworld, Cartpole, and Acrobot, respectively. Table 2: Zero-shot performance of non-equivariant models, equituned, and equizeroed models for LSTM on SCAN. LSTMs were trained for 200K iterations. We find that equizero outperforms other methods using non-equivariant pretrained models. Results are over three random seeds. (Table 2 includes Val. Acc.) |
| Hardware Specification | Yes | We use batch size 32 on a single Nvidia A100 GPU. |
| Software Dependencies | No | The paper mentions using GPT2 and refers to Stable-Baselines3 (via Raffin et al., 2021) but does not provide specific version numbers for software dependencies (e.g., PyTorch, TensorFlow, Python versions). |
| Experiment Setup | Yes | We first pretrain Deep Q-learning nets (DQNs) for each of the Gridworld, Cartpole, and Acrobot environments using the default architecture from Raffin et al. (2021) with 103k parameters. We pretrained all the models using a learning rate 10 4. We used training time steps as 100k, 100k, and 70k for Gridworld, Cartpole, and Acrobot, respectively. For pretraining, we use a learning rate of 10 4, whereas for finetuning, we used learning rates 10 5 and 2 10 5 for Add Jump and Around Right, respectively. |