Deconfounded Representation Similarity for Comparison of Neural Networks
Authors: Tianyu Cui, Yogesh Kumar, Pekka Marttinen, Samuel Kaski
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show that deconfounding the similarity metrics increases the resolution of detecting functionally similar neural networks across domains. Moreover, in real-world applications, deconfounding improves the consistency between CKA and domain similarity in transfer learning, and increases correlation between CKA and model out-of-distribution accuracy similarity. |
| Researcher Affiliation | Academia | Tianyu Cui Department of Computer Science Aalto University tianyu.cui@aalto.fi Yogesh Kumar Department of Computer Science Aalto University yogesh.kumar@aalto.fi Pekka Marttinen Department of Computer Science Aalto University pekka.marttinen@aalto.fi Samuel Kaski Department of Computer Science Aalto University and University of Manchester samuel.kaski@aalto.fi |
| Pseudocode | No | The paper describes algorithms and methods in prose and mathematical equations but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] We provide the code in the supplemental material. |
| Open Datasets | Yes | Setup: We check if similarity measures can identify functionally similar NN representations from random NN representations. For each model block of Res Nets (containing 2-3 convolutional layers), we generate two distributions of similarities: the null distribution H0 and the alternative distribution H1. The H0 contains similarities between 50 pairs of random Res Nets on CIFAR-10 test set. ... Distribution H1 contains similarities between the pretrained Image Net NN and each of the 50 Res Nets trained on CIFAR-10 from scratch with different random initializations, on the same CIFAR-10 test as H0. |
| Dataset Splits | Yes | We compute the layer-wise CKA and d CKA between each FT model and the corresponding PT model on the test set of the target domain [22]. ... 2. Evaluate the OOD accuracy of each model on CIFAR-10-C [40], acc(fi), and select the most accurate Res Net as the reference model, f ; 3. Compute the similarity between each fi and f , s(fi, f ), of each block on CIFAR-10 test set (in-distribution similarity)... |
| Hardware Specification | Yes | In experiments, computing d CKA between two XLM-Ro BERTa models [38] takes 0.37 ± 0.11s longer than CKA for each layer on 3000 random English sentences with a single 2080Ti GPU. |
| Software Dependencies | No | The paper mentions using specific models like Res Nets, XLM-RoBERTa, EfficientNet-B0, and Distil-RoBERTa, and refers to PyTorch models in a footnote, but it does not specify exact version numbers for any software libraries, frameworks, or programming languages used. |
| Experiment Setup | Yes | Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] See Appendix C. |