Neural Markov Jump Processes
Authors: Patrick Seifner, Ramses J Sanchez
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We test our approach on synthetic data sampled from ground-truth Markov jump processes, experimental switching ion channel data and molecular dynamics simulations. |
| Researcher Affiliation | Academia | 1Lamarr Institute, Bonn, Germany. 2University of Bonn, Bonn, Germany. Correspondence to: Patrick Seifner <seifner@cs.uni-bonn.de>, Rams es J. S anchez <sanchez@bit.uni-bonn.de>. |
| Pseudocode | Yes | Algorithm 1. Training Neural MJP |
| Open Source Code | Yes | Source code to reproduce our experiments is available online.1 1https://github.com/pseifner/Neural MJP |
| Open Datasets | No | For Ion Channel Data: 'The dataset was made available to us via private communication'. For Alanine Dipeptide: 'This simulation dataset was originally used by Wang et al. (2019) and was made available to us via private communication'. For Lotka-Volterra: 'The dataset was made available to us via private communication'. |
| Dataset Splits | Yes | Out of the 5000 time series, we use 70 batches of size 64 for training and leave the remaining time series for test and validation." and "We use 100 time series of length 200 for prediction and 10 time series of length 100 for test and validation each. |
| Hardware Specification | Yes | Each experiment was performed on a single NVIDIA Ge Force GTX 1080 Ti. |
| Software Dependencies | No | The paper mentions 'torchdiffeq python package (Chen, 2018)' but does not provide explicit version numbers (e.g., 1.0, 2.1) for it or any other key software dependencies. |
| Experiment Setup | Yes | To specify the posterior model, we fix the dimension H of h T to 256 in all experiments. Its ODE-RNN encoder uses a GRU network (Cho et al., 2014), with a hidden dimension of 256, for the instantaneous updates, and solves a Neural ODE network, parametrized by an MLP with [256, 256] layers, backwards in time with the Runge-Kutta method, starting from T with initial condition h0 = 0. The functions Λψ and Ψϕ are both MLPs with [128, 128] and [256, 256, 128] internal layers, respectively. The posterior master equation (Eq. 4) is solved with the adaptive step Dormand Prince method. The Kullback-Leibler divergence in Eq. 5 is approximated via Gaussian quadrature with 200 points. Next, and regarding the generative prior, we set the dimension p of the random input vector ε to 64, and model Φθ with a single hidden layer of 64 units. The emission model, when used, is set to a Gaussian model, whose mean and variance are modelled with and MLP of [128, 128] layers. Finally, all parameters are optimized using Adam (Kingma & Ba, 2015), with learning rates varying from 1 10 3 to 1 10 4, and possible learning rate annealing. |