Learning diffusion at lightspeed
Authors: Antonio Terpin, Nicolas Lanzetti, Martín Gadea, Florian Dorfler
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Via exhaustive numerical experiments, we show that, in the case of potential energies only, JKOnet outperforms the state-of-the-art in terms of solution quality, scalability, and computational efficiency and, in the until now unsolved case of general energy functionals, allows us to also learn interaction and internal energies that explain the observed population trajectories. |
| Researcher Affiliation | Academia | Antonio Terpin ETH Zürich aterpin@ethz.ch Nicolas Lanzetti ETH Zürich lnicolas@ethz.ch Martín Gadea ETH Zürich mgadea@ethz.ch Florian Dörfler ETH Zürich dorfler@ethz.ch |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Source code: https://github.com/antonioterpin/jkonet-star |
| Open Datasets | Yes | We deploy JKOnet to analyze the embryoid body single-cell RNA sequencing (sc RNA-seq) data [35] describing the differentiation of human embryonic stem cells over a period of 27 days. |
| Dataset Splits | Yes | We train the time-varying extension of JKOnet V , JKOnet and JKOnet-vanilla for 100 epochs on 60% of the data at each time and we compute the EMD between the observed µt (40% remaining data) and one-step ahead prediction ˆµt at each timestep. |
| Hardware Specification | Yes | The empirical data was collected entirely on an Ubuntu 22.04 machine equipped with an AMD Ryzen Threadripper PRO 5995WX processor and a Nvidia RTX 4090 GPU. |
| Software Dependencies | No | The paper mentions software like JAX, POT library, OTT-JAX library, and SciPy package, but does not provide specific version numbers for these dependencies. |
| Experiment Setup | Yes | We use the Adam optimizer [26] with the parameters β1 = 0.9, β2 = 0.999, ε = 1e-8, and constant learning rate lr = 1e-3. The model is trained with gradient clipping with maximum global norm for an update of 10. We process the data in batches of 250. The neural networks of potential and interaction energies are multi-layer perceptrons with 2 hidden layers of size 64 with softplus activation functions and a one-dimensional output layer (cf. [9]). |