Microsoft Says Its New Computer Vision System Can Outperform Humans

Recently, the researchers of Microsoft claims in published paper that their team has developed the world’s first computer machine, which is capable to outperform humans on certain popular benchmark. According to estimation of experts, humans can identify or classify images at ImageNet dataset and in this process the error rate will be 5.1 per cent, whereas; in recent published paper team of researchers indicated that due to deep-learning based system, successfully they have achieved an error rate of 4.94 per cent, which is little less than human error rate.

Microsoft’s paper was published in less than a month after Baidu’s paper. It is well known thing that Baidu is most famous search engine in China as Google is ban in China. Baidu published its paper indicating its record-setting system that claimed an error rate of 5.98 per cent with the help of homemade supercomputing architecture. The performance was based on the competition on ImageNet, which belongs from the team of Google researchers, who have also built a deep learning system with an error rate of 6.66 in 2014.

The paper of Microsoft states, “As per our knowledge, we are the first who have published instance of surpassing humans with the help of visual recognition challenge, whereas; on the native side, our algorithms are still making mistakes in certain cases which are not so difficult for humans, especially then when it requires a high-level knowledge or context understanding”.

Whereas; our all algorithms have given a superior result with this specific dataset, which does not indicate that machine’s vision can outperforms the human vision on object recognition in general. At the end of research, we believe that our results are showcasing the tremendous potential of system algorithms to match the performance level with human for certain visual recognition tasks.

According to Jian Sun, who is researcher in Microsoft, the difference is plainer English in a blog post of Microsoft; “Humans can easily make the difference between a sheep and a cow or between other things whereas; computers are not perfect in these simple tasks. However, when it comes to differentiate between breeds of sheep’s than here outperform humans and it proves that computer can be trained to look at the certain details such as; size, shape, texture, and context of the image to identify the difference that can’t be observed by humans.

As human or researcher if you have curiosity to learn how deep learning works, why in these days it’s a hot area, how it can be used in commercial applications, then you should attend the Structure Data conference of Microsoft at New York between March 18 and 19. Speakers will be machine learning and deep learning experts from Microsoft, Facebook, Hampton Creek, Yahoo, Spotify, NASA and Stanford as well as also from startups such as; Enlitic, MetaMind, Blue River Technology and TeraDeep. Later in Structure Intelligence conference, which will be organized between Sept. 22 and 23 in San Francisco, we will dive even deeper into artificial intelligence, says spokesperson from Microsoft