Neural Lagrangian Schr\"{o}dinger Bridge: Diffusion Modeling for Population Dynamics

Authors: Takeshi Koshizuka, Issei Sato

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that the proposed method can efficiently approximate the population-level dynamics even for high-dimensional data and that using the prior knowledge introduced by the Lagrangian enables us to estimate the sample-level dynamics with stochastic behavior.
Researcher Affiliation Academia Takeshi Koshizuka, Issei Sato The University of Tokyo {koshizuka-takeshi938444, sato}@g.ecc.u-tokyo.ac.jp
Pseudocode Yes Algorithm 1 Training of NLSB
Open Source Code Yes Our code is available at https://github.com/take-koshizuka/nlsb.
Open Datasets Yes We evaluated on embryoid body sc RNA-seq data (Moon et al., 2019). The sc RNA-seq data are licensed under Creative Commons Attribution 4.0 International license.
Dataset Splits Yes We generated 2048 and 512 samples for each time point as training and validation data, respectively. We split the dataset into train, validation( 8.5%) and test data ( 15%).
Hardware Specification Yes Our experimental environment consists of an Intel Xeon Plantinum 8360Y (36-core) CPU and a single NVIDIA A100 GPU.
Software Dependencies No It mentions using 'torchdiffeq library' and 'torchsde library', but specific version numbers for these libraries or Python are not provided.
Experiment Setup Yes We trained all models with a batch size of 512 for each time point. The tuned weight coefficients are shown in Table 4. For the potential model Φθ(x, t), we set the number of Res Net layers M = 2, the step size h = 1.0, the rank of matrix rank(A) = 10, and the dimension of the hidden vector z to 2. In training all models, we used Adam optimizer to optimize all learnable parameters with a learning rate of 0.001 and the decay rate of β1 = 0.9, β2 = 0.999.