Unifying Likelihood-free Inference with Black-box Optimization and Beyond
Authors: Dinghuai Zhang, Jie Fu, Yoshua Bengio, Aaron Courville
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on sequence design application illustrate the benefits of the proposed methodology. |
| Researcher Affiliation | Academia | Dinghuai Zhang1,2 , Jie Fu1,2 , Yoshua Bengio1,2,3, Aaron Courville1,2,3 1Mila, 2University of Montreal, 3CIFAR Fellow Montreal, Canada |
| Pseudocode | Yes | Algorithm 1 SMC-ABC, Algorithm 2 FB-VAE, Algorithm 3 Sequential Neural Posterior, Algorithm 4 Design by Adaptive Sampling, Algorithm 5 Sequential Neural Likelihood, Algorithm 6 Iterative Scoring, Algorithm 7 Sequential Neural Ratio, Algorithm 8 Iterative Ratio, Algorithm 9 Iterative Posterior Scoring, Algorithm 10 Iterative Posterior Ratio |
| Open Source Code | No | The paper does not provide an explicit statement or link for the open-source code of their methodology. |
| Open Datasets | Yes | In this section, we systematically evaluate the proposed methods and baselines on four different in-silico biological sequence design benchmarks. ... Transcription factor binding sites (Tf Bind). ... In Barrera et al. (2016), the authors measure the binding properties... 5 untranslated regions (UTR). ... In Sample et al. (2019), the authors create a library... Antimicrobial peptides (AMP). We use the AMP dataset from (Witten & Witten, 2019)... Fluorescence proteins (Fluo). We use a pretrained model taken from Rao et al. (2019) to act as our task oracle... |
| Dataset Splits | Yes | For validation, we follow Angerm uller et al. (2020b) and use one task (ZNF200 S265Y R1) for hyperparameter selection. Then we test the algorithms performance on the other 14 held-out tasks. ... For validation, we follow (Angermuller et al., 2020b) and sweep each algorithm for fifty trials and pick the best configuration. We tune learning rate and whether to re-initialize the optimizer for each new round for all methods. We tune threshold for Db AS, FB-VAE and the methods that is with Example A. For the other choice of E, we tune the temperature. For evolution, we tune the number of offsprings for each sequence in generation list, the probability of substitution, insertion and deletion. For Tf Bind we use ZNF200 S265Y R1 for validation. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for the experiments, such as GPU/CPU models or memory specifications. |
| Software Dependencies | No | The paper mentions software like "VAE model", "Bi-directional long short-term memory model (Bi LSTM)", "Prot Albert", and "Python" but does not specify version numbers for these dependencies. |
| Experiment Setup | Yes | In every round, we allow each algorithm to query the black-box oracle for a batch of n sequences mi to obtain their true scores si, with n = 100 for all experiments. The total number of rounds differs across different tasks. ... Both the embedding dimension and the hidden size of LSTM is set to 32. For composite methods that use two models, we use one-layer LSTM for each of them. For the other algorithms that only use one LSTM, we set its number of layers to be two. ... We tune learning rate and whether to re-initialize the optimizer for each new round for all methods. We tune threshold for Db AS, FB-VAE and the methods that is with Example A. For the other choice of E, we tune the temperature. For evolution, we tune the number of offsprings for each sequence in generation list, the probability of substitution, insertion and deletion. |