Differentiable Segmentation of Sequences

Authors: Erik Scharwächter, Jonathan Lennartz, Emmanuel Müller

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The experiments show that our approach effectively solves all these tasks with a standard algorithm for gradient descent.
Researcher Affiliation Academia Erik Scharwächter TU Dortmund University, Germany Jonathan Lennartz University of Bonn, Germany Emmanuel Müller TU Dortmund University, Germany
Pseudocode Yes Algorithm 1 Data generating process of Arlot et al. (2019).
Open Source Code Yes Source codes for the model and all experiments can be found in the online supplementary material at https://github.com/diozaka/diffseg.
Open Datasets Yes We obtained official data for Germany from Robert Koch Institute.3 ... The data is publicly available under an open data license. ... We sample random sequences of length T = 1000 with 10 change points at predefined locations. ... Our experimental design exactly follows Arlot et al. (2019). ... We use the insect stream benchmark of Souza et al. (2020) for this purpose.
Dataset Splits No The paper describes training and evaluation on datasets and streams but does not specify explicit train/validation/test splits by percentages or counts, nor does it explicitly mention using a 'validation set'.
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU/GPU models, memory, or cloud instance types) used for running the experiments.
Software Dependencies No We implemented our model in Python1 using Py Torch2 and optimize the parameters with ADAM (Kingma & Ba, 2015). The paper mentions Python and PyTorch but does not specify their version numbers, nor any other library versions.
Experiment Setup Yes We set the window size to w = .5 and the power to n = 16. ... We perform training with ADAM with a learning rate of η = 0.01 for a total of 10000 training epochs. In the last 2048 epochs, we round the predictors ˆζt to the nearest integers to obtain a hard segmentation function ζ(t). We perform 10 restarts of the training procedure with random initialization and keep the model with the best fit for evaluation (highest log-likelihood).