A New Family of Generalization Bounds Using Samplewise Evaluated CMI

Authors: Fredrik Hellström, Giuseppe Durisi

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

Reproducibility Variable Result LLM Response
Research Type Experimental We now consider three deep learning settings and evaluate the bounds derived in Section 2. To the best of our knowledge, the tightest average generalization bounds available in the literature for typical deep learning scenarios, such as CNNs trained on MNIST [41] or CIFAR10 [42], are found in [17]. These numerical results are based on [17, Eq. 22], which is a version of (4) where the e-CMI is replaced by the f-CMI.
Researcher Affiliation Academia Fredrik Hellström Chalmers University of Technology Gothenburg, Sweden frehells@chalmers.se Giuseppe Durisi Chalmers University of Technology Gothenburg, Sweden durisi@chalmers.se
Pseudocode No The paper provides theoretical results in the form of theorems and lemmas, but does not include any pseudocode or algorithm blocks.
Open Source Code Yes The code for our experiments is largely based on the code from [17].3 Available: https://github.com/hrayrhar/f-CMI. (Footnote 3)
Open Datasets Yes First, we consider a CNN trained with Adam (a variant of SGD) on a binarized version of MNIST, where we only consider the digits 4 and 9. Next, in Figure 2b, we look at Res Net-50 pretrained on Image Net and fine-tuned using SGD on CIFAR10.
Dataset Splits Yes Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] This is specified in the supplementary material.
Hardware Specification Yes Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] This is specified in the supplementary material.
Software Dependencies No The paper mentions training algorithms like Adam and SGD but does not specify software versions for libraries or environments used for implementation (e.g., PyTorch version, Python version).
Experiment Setup Yes Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] This is specified in the supplementary material.