GeoPhy: Differentiable Phylogenetic Inference via Geometric Gradients of Tree Topologies

Authors: Takahiro Mimori, Michiaki Hamada

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

Reproducibility Variable Result LLM Response
Research Type Experimental In experiments using real benchmark datasets, Geo Phy significantly outperformed other approximate Bayesian methods that considered whole topologies.
Researcher Affiliation Academia Takahiro Mimori Waseda University RIKEN AIP takahiro.mimori@aoni.waseda.jp Michiaki Hamada Waseda University AIST-Waseda CBBD-OIL Nippon Medical School mhamada@waseda.jp
Pseudocode Yes Algorithm 1 Geo Phy algorithm
Open Source Code Yes Our implementation is found at https://github.com/m1m0r1/geophy
Open Datasets Yes We compared the marginal log-likelihood (MLL) estimates for the eight real datasets (DS1-8) [8, 6, 37, 9, 17, 43, 38, 30].
Dataset Splits No The paper does not explicitly describe train/validation/test splits for its own model training in the conventional supervised learning sense. It performs inference on the entire given datasets.
Hardware Specification Yes We measured CPU runtimes using a single thread on a 20-core Xeon processor.
Software Dependencies Yes we replicated the Mr Bayes SS runs using Mr Bayes version 3.2.7a.
Experiment Setup Yes In all of our experiments, we used Adam optimizer with an initial learning rate of 0.0001. The learning rate was then multiplied by 0.75 after every 200,000 steps. Similar to approaches taken by Zhang and Matsen IV [42], we incorporated an annealing procedure during the initial consumption of 100,000 MC samples. Specifically, we replaced the likelihood function in the lower bound with P(Y |Bτ, τ)β and linearly increased the inverse temperature β from 0.001 to 1 throughout the iterations.