Hyper-Representations: Self-Supervised Representation Learning on Neural Network Weights for Model Characteristic Prediction
Year: 2021
Type: preprint
Abstract: Datasets to NeurIPS 2021 accepted paper "Self-Supervised Representation Learning on Neural Network Weights for Model Characteristic Prediction". Datasets are pytorch files containing a dictionary with training, validation and test sets. Train, validation and test sets are custom dataset classes which inherit from the standard torch dataset class. Corresponding code an be found at https://github.com/HSG-AIML/NeurIPS_2021-Weight_Space_Learning. Datasets 41, 42, 43 and 44 are our dataset format wrapped around the zoos from Unterthiner et al, 2020 (https://github.com/google-research/google-research/tree/master/dnn_predict_accuracy) Abstract: Self-Supervised Learning (SSL) has been shown to learn useful and information-preserving representations. Neural Networks (NNs) are widely applied, yet their weight space is still not fully understood. Therefore, we propose to use SSL to learn neural representations of the weights of populations of NNs. To that end, we introduce domain specific data augmentations and an adapted attention architecture. Our empirical evaluation demonstrates that self-supervised representation learning in this domain is able to recover diverse NN model characteristics. Further, we show that the proposed learned representations outperform prior work for predicting hyper-parameters, test accuracy, and generalization gap as well as transfer to out-of-distribution settings.
Source: arXiv (Cornell University)
Institution: University of St.Gallen
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4285662467