Task Discovery: Finding the Tasks that Neural Networks Generalize on

Authors: Andrei Atanov, Andrei Filatov, Teresa Yeo, Ajay Sohmshetty, Amir Zamir

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

Reproducibility Variable Result LLM Response
Research Type Experimental We propose a task discovery framework that automatically finds examples of such tasks via optimizing a generalization-based quantity called agreement score. We demonstrate that one set of images can give rise to many tasks on which neural networks generalize well. These are the questions we address in this paper.
Researcher Affiliation Academia Andrei Atanov Andrei Filatov Teresa Yeo Ajay Sohmshetty Amir Zamir Swiss Federal Institute of Technology (EPFL)
Pseudocode No The paper includes a diagram illustrating the meta-optimization process (Fig. 3-left) but no structured pseudocode or algorithm blocks.
Open Source Code Yes https://taskdiscovery.epfl.ch (on page 1). Additionally, in Section A, item 3a states: '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]'
Open Datasets Yes We take the set of images X from the CIFAR-10 dataset [39]
Dataset Splits No The paper mentions splitting the original training set into 45K images for Xtr and 5K for Xte, referring to a train and test set, but does not specify a separate validation split.
Hardware Specification Yes This allows us to run the discovery process for the Res Net-18 model on the CIFAR-10 dataset using a single 40GB A100.
Software Dependencies No The paper mentions 'PyTorch [63]' and 'Adam [34] optimizer' but does not provide specific version numbers for these software components.
Experiment Setup Yes We use Res Net-18 [24] architecture and Adam [34] optimizer as the learning algorithm A, unless otherwise specified. We measure the AS by training two networks for 100 epochs, which is enough to achieve zero training error for all considered tasks. To make it feasible, we limit the number of inner-loop optimization steps to 50.