Local Intrinsic Dimensional Entropy

Authors: Rohan Ghosh, Mehul Motani

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments and Discussions Here we showcase two experiments, where we compute the ID-Entropy of the feature layers of classifier and autoencoder architectures on MNIST and CIFAR-10, and contrast it with generalization performance.
Researcher Affiliation Academia Rohan Ghosh1, Mehul Motani1,2 1 College of Design and Engineering, Department of Electrical and Computer Engineering, National University of Singapore 2 N.1 Institute for Health, Institute for Digital Medicine (Wis DM), Institute of Data Science, National University of Singapore
Pseudocode Yes Algorithm 1: Estimation of ID-Entropy Input: S = {X1, .., Xm} (i.i.d samples of RV X), a global ID estimator f ID(S ) of points in S , & parameters (k, n). Output: IDX (Estimate of ID-Entropy of X) 1: IDsum = 0 2: for j = 1, 2m/n , 3m/n , .. . . . , m do 3: Let S = k-nearest neighbors of Xj in S 4: IDsum = IDsum + f ID(S ) 5: IDX = IDsum/n
Open Source Code Yes code is available at: https://github.com/kentridgeai/ID-Entropy.
Open Datasets Yes Experiments and Discussions Here we showcase two experiments, where we compute the ID-Entropy of the feature layers of classifier and autoencoder architectures on MNIST and CIFAR-10
Dataset Splits No The paper mentions 'training data size' and 'test accuracy' for MNIST and CIFAR-10 but does not specify explicit training, validation, or test dataset splits (e.g., percentages or sample counts).
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python, TensorFlow, PyTorch versions).
Experiment Setup Yes For all experiments we estimated the ID-entropy using Algorithm 1, with n = 2000 and k = 100. We used Fisher’s intrinsic dimensionality estimator in (Bac and Zinovyev 2020) as the global ID estimator f ID, as we found it to be a robust choice. With a fixed 4-layer CNN architecture for MNIST and a Res Net-44 for CIFAR-10, we repeat the training routine with different choices of the training data size and random network initializations. For both MNIST (4-layer CNN) and CIFAR-10 (Res Net-44), we add label noise by randomly changing the label of a training data point with a certain probability.