Ant Group's 11 papers were successfully selected for ICLR 2024, the top international AI conference

Recently, ICLR 2024, the top artificial intelligence conference, announced the admission results. Ant Group had 11 papers accepted at this conference, of which 1 was rated as an oral report, 3 were selected as focus reports, and the other 7 were poster presentations. Ant Group’s progress in the artificial intelligence academic community has attracted much attention.

(Picture: Ant Group’s "Multi-granularity Noise Association Learning in Long Videos" was included as an Oral paper)

This year, the ICLR organizing committee received 7,262 paper submissions, with an acceptance rate of approximately 31%. According to the acceptance results, 1.2% of the papers will be accepted as Oral papers, and these authors will receive a 10-minute oral speech opportunity. Another 5% of papers are accepted as Spotlight papers, and these authors will have 4 minutes of spotlight time. The remaining papers will be presented in poster format. Overall, Oral papers have the highest importance, followed by Spotlight papers, and Poster papers have the lowest importance.

Every year, a considerable number of ICLR Oral papers will be rated as "ICLR Best Papers", which also means that they guide the research direction for the new year. This year, ICLR selected 85 Oral papers, including Ant Group's "Multi-granularity Correspondence Learning from Noisy Instructional Videos" (Multi-granularity Correspondence Learning from Noisy Instructional Videos). This paper proposes a method of learning using noisy teaching videos, which improves the performance and robustness of the model through associated learning at multiple granularities. This research is of great significance for solving the noise and uncertainty problems that exist in the real world, and provides new ideas for further development in the field of video understanding.

Short videos have become the main form of entertainment in people’s daily lives, and multi-modal technology is a popular research direction in the current AI field. However, due to the high computational resource overhead, existing video work mainly focuses on segment understanding, while ignoring the temporal dependencies in long videos. To solve this problem, this paper transforms long video learning into association alignment between short video clips. Aiming at the problem of noise correlation between video and text, the study proposed a unified optimal transmission alignment scheme. This scheme significantly improved the understanding of long videos and also saved time. Through this research, we can better understand long videos and be more accurate and efficient in processing the association between video and text.

This solution is very versatile, and the proposed noise correlation processing method is suitable for other model pre-training learning that requires content alignment.

Spotlight has included three papers, namely "iTransformer: Inverted Transformers Are Effective for Time Series Forecasting" (iTransformer: Inverted Transformers are more effective for time series forecasting), "Enhanced Face Recognition using Intra-class Incoherence Constraint" "(Face recognition technology enhanced by intra-class inconsistency constraints) and "Finite-State Autoregressive Entropy Coding for Efficient Learned Lossless Compression" (Learnable autoregressive model based on lookup table implementation for efficient lossless compression algorithm). The first paper introduces a new time series forecasting method, which achieves comprehensive leading results in complex time series forecasting tasks by breaking the conventional model structure. This research has important implications for improving the accuracy and efficiency of time series forecasting. The second paper introduces a new method to improve the accuracy of face recognition. This method utilizes intra-class inconsistency constraints to further optimize face recognition technology. This research is of great significance for improving the performance and accuracy of face recognition systems. The third paper proposes a learnable autoregressive model implemented based on lookup tables for efficient lossless compression. This research realizes a lossless compression algorithm with high compression rate and high throughput rate, which has important application value for data compression and storage. The publication of these three papers has made important breakthroughs and progress in their respective fields, providing strong support for research and applications in related fields. Their emergence has enriched the research results of academia and brought new possibilities to the development of related fields.

Since 2017, the number of papers received by ICLR every year has increased by 30%, and the two top artificial intelligence conferences, NeurIPS and ICML, have also shown rapid growth trends. At the recent NeurIPS conference, a total of 20 papers from Ant Group were included. These papers cover cutting-edge topics in computer vision, natural language processing, graph neural networks, image processing and other fields of artificial intelligence and machine learning. This achievement further proves Ant Group’s outstanding research strength and innovation capabilities in the field of artificial intelligence.

(Picture: ICLR’s annual number of papers since its establishment in 2013. Starting from 2017, the number of papers has increased. )

ICLR has been well received by the industry in recent years, mainly due to its Open Review review system. All submitted papers will be evaluated and questioned by all peers, and any scholar can evaluate papers anonymously or under their real name. After the public review is completed, the author of the paper can also adjust and modify the paper.

It is understood that in the past five years, Ant Group has published nearly 500 papers in top international academic journals and academic conferences, including more than 300 papers in the field of AI. Ant Group continues to invest in technology in the field of artificial intelligence. Based on the needs of large-scale business scenarios, it has laid out technical fields including large models, knowledge graphs, operations optimization, graph learning, and trusted AI.

The above is the detailed content of Ant Group's 11 papers were successfully selected for ICLR 2024, the top international AI conference. For more information, please follow other related articles on the PHP Chinese website!