Log-Euclidean Signatures for Intrinsic Distances Between Unaligned Datasets

Authors: Tal Shnitzer, Mikhail Yurochkin, Kristjan Greenewald, Justin M Solomon

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate properties of the LES distance on various data types, including 3-4D point clouds of geometric shapes, single cell RNA-sequencing data, and NN embeddings, where we show that the LES distance provides useful insights on architectures and embedding structures.
Researcher Affiliation Collaboration 1MIT CSAIL 2IBM Research 3MIT-IBM Watson AI Lab. Correspondence to: Tal Shnitzer <talsd@mit.edu>.
Pseudocode Yes Algorithm 2 summarizes our approach.
Open Source Code Yes The code is available in https://github.com/shnitzer/les-distance.
Open Datasets Yes We apply LES to data of reprogrammed mouse embryonic cells (Schiebinger et al., 2019)... We train Meta Opt Net-SVM (Lee et al., 2019) for benchmark datasets CIFAR-FS (Bertinetto et al., 2018) and FC100 (Oreshkin et al., 2018)...
Dataset Splits Yes Test accuracies (first row in Table 1) in each setting are computed using 1000 random test tasks, and we report mean and standard deviation. ... To evaluate correlations we first sample n = 10 tasks from the train task distribution. ... To obtain a range of new tasks from easier to more difficult ones, we consider mixtures of train and test classes: only test classes (most difficult), 4 test and 1 train classes, 3 test and 2 train classes, 2 test and 3 train classes, 1 test and 4 train classes, and only train classes (easiest).
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory specifications) used for running the experiments.
Software Dependencies No The paper mentions software like 'Gephi software' and 'Ripser.py' in the references but does not specify the version numbers of any core software dependencies used for their own methodology or experiments (e.g., Python version, specific libraries with versions).
Experiment Setup Yes Throughout this section we compute LES according to Algorithm 2 with σ2 = 2 median d2(xi, xj) , K {200, 500}, M = 2K and γ [10 8, 10 5]. The exact parameters used in each application are described in the appendices.