Learning Discrete Energy-based Models via Auxiliary-variable Local Exploration

Authors: Hanjun Dai, Rishabh Singh, Bo Dai, Charles Sutton, Dale Schuurmans

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We experimentally evaluated the approach on both synthetic and real-world tasks. For a program synthesis problem, we observe significant accuracy improvements over the baseline methods. More notably, for a software testing task, a fuzz test guided by an EBM achieves comparable performance to a well-engineered fuzzing engine on several open source software projects.
Researcher Affiliation Industry Hanjun Dai, Rishabh Singh, Bo Dai, Charles Sutton, Dale Schuurmans Google Research, Brain Team {hadai, rising, bodai, charlessutton, schuurmans}@google.com
Pseudocode Yes Algorithm 1: Main algorithm of ALOE; Algorithm 2: Update sampler q
Open Source Code Yes Please also refer to our open sourced implementation for more details 1. 1https://github.com/google-research/google-research/tree/master/aloe
Open Datasets Yes We first collect synthetic 2D data in a continuous space [60]... We collect three software binaries (namely libpng, libmpeg2 and openjpeg) from OSS-Fuzz3 as test target... Data is generated synthetically, following Devlin et al. [3].
Dataset Splits No The paper mentions "10k test examples" and "training data is generated on the fly" but does not provide specific training/validation/test splits or mention a validation set.
Hardware Specification No The paper does not specify any hardware details such as GPU or CPU models used for experiments.
Software Dependencies No The paper does not specify software versions for any dependencies (e.g., Python, TensorFlow, PyTorch) used for implementation.
Experiment Setup Yes All the methods learn the same score function f, which is parameterized by a 4-layer MLP with ELU [61] activations and hidden layer size= 256. ... we train a truncated EBM with a window size of 64. ... We train baseline seq2seq with 483M examples (denoted as seq2seq-init), and fine-tune with additional 264M samples (denoted as seq2seq-tune) with reduced learning rate.