Topics

We invite submissions studying the effect of algorithms on the training time and performance of neural networks. Possible topics for submissions include, but are not limited to:

  • Novel algorithms for training deep neural networks. These methods may be in the form of optimizers, data sampling strategies, models, hyperparameter tuning methods, training tricks, etc.

  • Investigations of existing methods. For example, these investigations could shed light on the relationship between existing methods, show in which situations specific algorithms are beneficial, disentangle multiple effects of a method, or highlight failure cases of existing approaches.

  • Metrics for measuring algorithms in deep learning. How can we best measure algorithmic improvements for neural network training? We invite papers that present benchmarks and metrics to provide a clearer signal of improvement in research on new deep learning methods.

Providing state-of-the-art results is not a requirement for acceptance. However, this workshop emphasizes a fair comparison between methods. To facilitate this, we strongly encourage providing all algorithmic details (e.g., in the form of code) so that results can be easily reproduced and verified by third parties. For novel training algorithms, we recommend that the workshop papers also submit to the MLCommons™ Algorithmic Efficiency benchmark but this is not a prerequisite for submitting.

Submission

Papers have to be submitted via OpenReview:

openreview.net/group?id=NeurIPS.cc/2022/Workshop/HITY.

The review process is double-blind.

Submissions must be anonymous, follow the NeurIPS 2022 style guide (LaTex style files, Overleaf), and should be no more than 4 pages excluding references, acknowledgments, and supplementary materials. Reviewers are not required to read the supplementary materials. Adding the NeurIPS checklist is recommended but not mandatory.

Accepted submissions will be presented in a poster session. Additionally, they will be made publically available as non-archival reports, allowing for subsequent publication in an archival conference or journal.

We will not accept submissions that have already been accepted for publication in other venues with archival proceedings (including publications that will be presented at the NeurIPS main conference). We discourage dual submission to concurrent NeurIPS workshops. Extended abstracts of papers under review at other conferences/journals can be submitted if this is ok for the conference/journal in question (if in doubt, please check with them first).