Microsoft’s Deep Learning Surpasses Humans and Google’s GoogLeNet

AI may become a driving force of our whole civilization, quietly running processes behind everything we do.

A group of researchers at Microsoft Research branch in Asia published an academic paper on recent results they achieved with convolutional neural networks. They claim that their system surpassed humans in image recognition and classification task on the 1000-class Image Net 2012 data set.

Their mean machine achieved high level of image recognition accuracy with only 4,94 % error rate, compared with a 5.1 % of humans and a 6,66 % of Google’s famous GoogLeNet, that used to be a record holder in a computer vision and pattern recognition until now.

The neural network is one of the approaches to machine learning. It is loosely based on the model of how the human brain works. It replicates connections which are created between neurons when the brain performs certain operations. During the learning process, some of these connections, depending on the relevance of information, will grow stronger, while those processing less relevant information weaken. In machine learning these connections are formed on statistical regularities and patterns drawn from processing big data from which computer develops self-learning algorithms.

Back in 1997, the entire world became aware of the great potential of the artificial intelligence (AI), when IBM’s supercomputer, known as Deep Blue, beat former World Chess Champion Garry Kasparov. From that moment, all big software companies began frantically pursuing the idea of building a self – learning machine that could outperform human brain power. Mega companies, like Microsoft, Google, Facebook and IBM are at the forefront of this research, which basically promises to create a self-thought machine, capable of mastering difficult cognitive tasks, like understanding natural language and human-like visual recognition.
According to Wired, investments in the AI sector have been continuously expanding over 60% on average in the past four years, a trend that is likely to continue. The number of startups is hoping to cash-in on this hype, developing algorithms and apps that could find its practical use on some AI platforms.

Experts rarely get into predictions about the future of machine learning beyond the five-year plan, which will mainly be focused on further upgrades of visual and speech recognition and augmenting capacities of search engines, areas in which machines can be thought by feeding them a large amount of data. But, there are many more promising fields in which refined processing may yield some groundbreaking results with a potential to completely change the global economy, medical diagnostics, transportation services, teaching, and who knows what else. From this standpoint, it is hard to accurately predict what can happen beyond the five-year time-frame. AI may become a driving force of our whole civilization, quietly running processes behind everything we do. In one way or another, it will transform many jobs in the future, even completely replace humans in some of them, but it will probably never become the kind of self-aware, omnipresent and omnipotent entity that many people are afraid of these days. In fact, each branch of AI’s application will most likely be so specialized, just as today’s technology is, that such an outcome is highly unlikely. Smart machines will be able to outperform humans in one particular job or a couple of related skills, but it will still be an autistic alien, maniacally pursuing prescribed tasks, as it is supposed to be. However disturbing it may sound, we should prefer these entities be more alien than human-like. As long as it stays that way, we are on the safe side.

That AI doesn’t pose any real threat to humanity in the near future is an opinion shared almost unanimously among participants at the 2015 Deep Learning Summit in San Francisco, and explicitly confirmed by Andrew Ng, Associate Professor at Stanford and the Chief Scientist at Baidu, considered as one of the leading researchers in the field of deep learning.