TWC issue #5

SOTA updates between 29 Aug–4 Sept 2022
- Temporal action localization
- Information retrieval
- Link prediction
This post is a consolidation of daily twitter posts tracking SOTA changes.
Official code release (with pre-trained models in most cases) also available for these tasks
#1 SOTA in temporal action localization on THUMOS’14 dataset

Paper: ActionFormer: Localizing Moments of Actions with Transformers
Submitted on 16 Feb 2022 (v1), last revised 28 Aug 2022 (v2) . Code updated 31 Aug 2022
Code link: Github link released by: Yin Li (last author in paper) Model link: Pretrained models on Github page
Notes: This model combines a multiscale feature representation with local self-attention, and uses a light-weighted decoder to classify every moment in time and estimate the corresponding action boundaries
Model Name: ActionFormer (I3D features)
Score (↑) : 66.8(Prev: 56.7)
Δ: 10.1 (Metric: mAP)
License: MIT license
Demo page link? None to date
Google colab link? None to date
Container image? None to date
#1 SOTA in Information retrieval on CQADupStack dataset


Paper: SGPT: GPT Sentence Embeddings for Semantic Search
Submitted on 17 Feb 2022 (v1), last revised 28 Aug 2022 (v5) . Code updated 31 Aug 2022
Code link: Github link released by: Muennighoff (last author in paper) Model link: Pretrained models can be accessed from HuggingFace. Code examples for access on Github page
Notes: Large language models do not perform well on embedding similarity or semantic search task. This paper suggests an approach to leverage large language models for these two tasks. Two architectures are proposed - one for embeddings and the other for semantic search.
Model Name: SGPT-BE-5.8B
Score (↑) : .16 (Prev: .145)
Δ: 015 (Metric: mAP@100 )
License: MIT license
Demo page link? This app compares SOTA models for sentence similarity
Google colab link? None to date
Container image? None to date
#1 SOTA in link prediction on OGBL-DDI dataset

Paper: ADAPTIVE GRAPH DIFFUSION NETWORKS
Submitted on 30 Dec 2020 (v1), last revised 02 Sept 2022 (v2) . Code updated 21 Sept 2021. Code has not been updated after the latest paper revision
Code link: Github link released by: Chuxiong Sun (first author in paper) Model link: Trained models not released yet
Notes: Standard graph diffusion methods combine large and dense powers of the transition matrix with predefined weighting coefficients. Instead, this paper combines smaller multi-hop node representations with learnable and generalized weighting coefficients. This approach is claimed to address the over-fitting and over-smoothing problems of deep GNNS.
Model Name: AGDN
Score (↑) : .9538 (Prev: .9284)
Δ: 015 .0254 (Metric: Test hits@20)
License: MIT license
Demo page link? None to date
Google colab link? None to date
Container image? None to date