NeurIPS 2022: My Top Two ‘Practically-Relevant’ Papers from the Outstanding 13

[siteorigin_widget class=”thinkup_builder_divider”][/siteorigin_widget]

NeurIPS 2022 declared 13 submissions as outstanding papers from its main track.

  1. Is Out-of-distribution Detection Learnable?
  2. Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding
  3. Elucidating the Design Space of Diffusion-Based Generative Models
  4. ProcTHOR: Large-Scale Embodied AI Using Procedural Generation
  5. Using Natural Language and Program Abstractions to Instill Human Inductive Biases in Machines
  6. A Neural Corpus Indexer for Document Retrieval
  7. High-dimensional Limit Theorems for SGD: Effective Dynamics and Critical Scaling
  8. Gradient Descent: The Ultimate Optimizer
  9. Riemannian Score-Based Generative Modelling
  10. Gradient Estimation with Discrete Stein Operators
  11. An Empirical Analysis of Compute-optimal Large Language Model Training
  12. Beyond Neural Scaling Laws: Beating Power Law Scaling via Data Pruning
  13. On-Demand Sampling: Learning Optimally from Multiple Distributions

Here are my top two picks from this list. The criteria for consideration are perceived utility, and ease of application in real-life scenarios.

[1] A Neural Corpus Indexer for Document Retrieval

Why do I consider it important? Index-Retrieve is arguably the most widely-followed approach in document retrieval. This paper demonstrates how a sequence-to-sequence network called Neural Corpus Indexer (NCI) unifies training and indexing to significantly improve upon conventional systems, particularly those based on Inverted Index and Dense Retrieval approaches. A key aspect of this paper is to leverage a novel decoder architecture called Prefix-Aware Weight-Adaptive (PAWA) to generate document identifiers.

[2] Gradient Descent: The Ultimate Optimizer

Why do I consider it important? This paper demonstrates how to automatically compute hypergradients through the backpropagation approach. The technique published by the researchers enables gradient descent-based optimizers to tune hyperparameters through automatic differentiation.

Share this article.