ExPT: Synthetic Pretraining for Few-Shot Experimental Design

Authors: Tung Nguyen, Sudhanshu Agrawal, Aditya Grover

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate Ex PT on few-shot experimental design in challenging domains and demonstrate its superior generality and performance compared to existing methods.
Researcher Affiliation Academia University of California, Los Angeles {tungnd,adityag}@cs.ucla.edu, sudhanshuagr27@g.ucla.edu
Pseudocode No The paper includes architectural diagrams (Figure 2) but no structured pseudocode or algorithm blocks.
Open Source Code Yes The source code is available at https://github.com/tung-nd/Ex PT.git.
Open Datasets Yes We consider 4 tasks from Design-Bench1 [58]. For each task, Design-Bench provides a public dataset, a larger hidden dataset which is used to normalize the scores, and an oracle.
Dataset Splits No The paper describes few-shot settings and data usage for pretraining and adaptation, but does not explicitly define or specify a separate validation dataset split for hyperparameter tuning or model selection in a reproducible manner.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory specifications) are provided for the experimental setup.
Software Dependencies No No specific software dependencies with version numbers (e.g., PyTorch 1.9, Python 3.8) are provided.
Experiment Setup Yes For each domain, we pretrain Ex PT for 10,000 iterations with 128 synthetic functions in each iteration...To increase the diversity of synthetic data, we randomize the two hyperparameters, length scale U[5.0, 10.0] and function scale σ U[1.0, 1.0], when generating each function. Additionally, we add Gaussian noises N(0, 0.1) to each input x sampled from Dunlabeled...For each generated function, we use 100 points as context points and the remaining 128 as target points, and train the model to optimize (2).