当前位置：网站首页>AI training speed breaks Moore's law; Song shuran's team won the RSS 2022 Best Paper Award

AI training speed breaks Moore's law; Song shuran's team won the RSS 2022 Best Paper Award

2022-07-01 18:51:00 【Zhiyuan community】

from ：AI Technology Review

newest MLPerf The benchmark indicates ：AI The training speed of is almost twice that of last year

6 month 29 Japan , Open Engineering Alliance MLCommons Released MLPerf The latest training results of the benchmark , It is found that the training speed of machine learning system this year is almost twice that of last year , Beyond Moore's law （ Every time 18-24 Double in months ）.

MLPerf It consists of eight benchmarks ： Image recognition 、 Medical image segmentation 、 Two versions of object detection 、 speech recognition 、 natural language processing 、 Recommend and reinforce learning . In these eight benchmarks , NVIDIA's accelerators are all at the top , Excellent performance .

MLPerf Training v2.0 Results include from 21 Of different submitters 250 Multiple performance results , Include Azure、 Baidu 、 Dell 、 Fujitsu 、 Jijia 、 Google 、Graphcore、HPE、 wave 、 Intel -HabanaLabs、 lenovo 、Nettrix、NVIDIA、 Samsung and Supermicro wait .（IEEE Spectrum）

Song shuran won the robot top meeting RSS 2022 Announce the best Paper Award

The robot will RSS（Robotics: Science and Systems） On 6 month 27 solstice 7 month 1 Held in New York on the th , And published the best paper 、 Best system paper 、 All awards including best student thesis .

among , Song shuran, a Chinese scholar currently working at Columbia University, and his team won RSS 2022 Best Paper Award , The award-winning work is “Iterative Residual Policy for Goal-Conditioned Manipulation of Deformable Objects”. Song shuran is a well-known young scholar in the field of robot research , He has won many conference Best Paper Awards , And get 2022 Nian Suyou “ Nobel vane ” The Sloan research award .

Google proposes a new language model Mineva To solve the problem of quantitative reasoning

lately , Google Research Institute has proposed a language model that can solve mathematical and scientific problems with step-by-step reasoning ——Mineva.

According to the team , At present, language models are used in quantitative reasoning （ That is to combine mathematics and information to solve real-world problems ） The performance of is still far lower than that of human , They collect training data related to quantitative reasoning problems 、 Large scale training model and advanced reasoning technology , Realized AI Significant progress of the model in quantitative reasoning tasks .

Thesis link ： https://storage.googleapis.com/minerva-paper/minerva_paper.pdf

In terms of technical architecture ,Mineva Based on what Google proposed a few months ago PaLM framework . In training , Yes arXiv On 118G Scientific paper data set and contains LaTex The webpage of mathematical expression is trained . Studies have shown that ,Mineva In the high school mathematics competition level question bank MATH、 University level STEM Good performance on the task .

British scholars jointly released a 100 page survey report on causal machine learning

lately , Oxford University and University College London （UCL） A survey of causal machine learning has been jointly released by scholars of , Long 165 page . Causal machine learning mainly studies how to transform the data generation process into a structural causal model （SCM）, People can infer the changes in this process （ Intervention ） Influence , And what will happen afterwards （ The counterfactual ）.

This review discusses five major directions of causal machine learning ： Causal supervised learning 、 Causal generation modeling 、 Causal interpretation 、 Causal justice and causal reinforcement learning .