Scaling up Continuous-Time Markov Chains Helps Resolve Underspecification
Authors: Alkis Gotovos, Rebekka Burkholz, John Quackenbush, Stefanie Jegelka
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the effectiveness of our approach on synthetic and real cancer data. and In experiments on real cancer data, we demonstrate how some previous results may have been artifacts of underspecification, and how scaling up the analysis can result in more robust models of cancer progression. |
| Researcher Affiliation | Academia | Alkis Gotovos MIT alkisg@mit.edu Rebekka Burkholz Harvard University rburkholz@hsph.harvard.edu John Quackenbush Harvard University johnq@hsph.harvard.edu Stefanie Jegelka MIT stefje@mit.edu |
| Pseudocode | Yes | Algorithm 1: Sampling a set from the marginal CTMC |
| Open Source Code | Yes | The code used to run our experiments can be found at https://github.com/3lectrologos/time/ tree/clean. |
| Open Datasets | Yes | We next evaluate our approach on a data set that contains N = 378 tumor samples of glioblastoma multiforme, an aggressive type of brain cancer. The data is part of the TCGA Pan Cancer Atlas project (https://www.cancer.gov/tcga), and we obtained a preprocessed version via c Bio Portal (Cerami et al., 2012). |
| Dataset Splits | No | The paper does not provide specific dataset split information (exact percentages, sample counts, or detailed splitting methodology) for training, validation, or testing. |
| Hardware Specification | Yes | Table 1: Learning run times (Intel Core i9 CPU) |
| Software Dependencies | No | The paper mentions optimization methods like 'proximal Ada Grad method' and initial step sizes, but does not specify software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | The following experimental setup is common to all our experiments. To optimize the objective (3), we use a proximal Ada Grad method (Duchi et al., 2011). We fix the initial step size to η = 1, and the regularization weight to λ = 0.01. To initialize the parameter matrix Θ, we train a diagonal model for 50 epochs, and then draw each off-diagonal entry from Unif([ 0.2, 0.2]). For the gradient approximation (7), we use M = 50 samples in addition to 10 burn-in samples that are discarded. |