Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Intrinsic Dimension, Persistent Homology and Generalization in Neural Networks
Authors: Tolga Birdal, Aaron Lou, Leonidas J. Guibas, Umut Simsekli
NeurIPS 2021 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments show that the proposed approach can efficiently compute a network s intrinsic dimension in a variety of settings, which is predictive of the generalization error. This section presents our experimental results in two parts: (i) analyzing and quantifying generalization in practical deep networks on real data, (ii) ablation studies on a random diffusion process. |
| Researcher Affiliation | Academia | Tolga Birdal Stanford University EMAIL Aaron Lou Stanford University EMAIL Leonidas Guibas Stanford University EMAIL Umut Sim sekli INRIA & ENS PSL Research University EMAIL |
| Pseudocode | Yes | Algorithm 1: Computation of dim PH. 1 input :The set of iterates W = {wi}K i=1, smallest sample size nmin, and a skip step , α 2 output :dim PHW 3 n nmin, E [] 4 while n K do 5 Wn sample(W, n) // random sampling 6 Wn VR(Wn) // Vietoris-Rips filtration 7 E[i] Eα(Wn) P γ PH0(Wn) |I(γ)|α // compute lifetime sums from PH 9 m, b fitline (log(nmin : : K), log(E)) // power law on Ei 1(W) 10 dim PHW α 1 m |
| Open Source Code | Yes | To foster further developments at the the intersection of persistent homology and statistical learning theory, we release our source code under: https://github.com/tolgabirdal/PHDim_Generalization. |
| Open Datasets | Yes | In particular, we train Alex Net [KSH12], a 5-layer (fcn-5) and 7-layer (fcn-7) fully connected networks, and a 9-layer convolutional netowork (cnn-9) on MNIST, CIFAR10 and CIFAR100 datasets for multiple batch sizes and learning rates until convergence. |
| Dataset Splits | No | The paper states it trains models until convergence and measures generalization as the gap between training and test accuracies, but does not explicitly detail training, validation, and test dataset splits. |
| Hardware Specification | No | The paper mentions using PyTorch for implementation but does not provide specific details about the hardware (e.g., GPU model, CPU, memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions using PyTorch [PGM+19] and the associated persistent homology package torchph [CHU17, CHN19] but does not specify their version numbers. |
| Experiment Setup | Yes | In particular, we train Alex Net [KSH12], a 5-layer (fcn-5) and 7-layer (fcn-7) fully connected networks, and a 9-layer convolutional netowork (cnn-9) on MNIST, CIFAR10 and CIFAR100 datasets for multiple batch sizes and learning rates until convergence. For Alex Net, we consider 1000 iterates prior to convergence and, for the others, we only consider 200. ... We train a Lenet-5 network [LBBH98] on Cifar10 [Kri09] and compare a clean trianing with a training with our topological regularizer with λ set to 1. We train for 200 epochs with a batch size of 128 and report the train and test accuracies in Fig. 5 over a variety of learning rates. |