FOLLOW

FOLLOW

SHARE

An Introduction to the AI Supercomputers

'Who?', 'What?' and 'What's Next?' for the worlds biggest artificial intelligence supercomputers

28Jan

The Artificial Intelligence revolution is here. We are moving further into an age, where the imagination stirred from our childhood spent watching movies, is now becoming reality.

Leading us into this age are the typical (and untypical) tech giants, who are fiercely competing for the next break through.

The 'AI Supercomputer' is the next break through. In the world of AI, this is the equivalent of the US and USSR competing to put their guy on the moon first. Here is a profile of some of the giants locked into the AI space race...

Microsoft Project Oxford

What is it?

Project Oxford is Microsoft’s venture into the world of artificial intelligence and deep learning. It takes in several key areas, including image, facial, text and speech recognition, and hopes to implement the technology into its computer operating systems and smartphone software.

What can it do?

Oxford’s most recent release is a facial recognition algorithm that works in conjunction with Windows phone in an alarm feature called Mimicker Alarm. The alarm only turns off when the person using it shows the correct facial expressions, and there are plans for other mini-games to be used as well.

What can’t it do?

Project Oxford’s purpose is quite commercially-centric, with the end goal more focussed around boosting the sales of their smartphones and computer operating systems. The datasets it is dealing with are not as large as those dealt with by some of its competitors, and it requires a large amount of human input to tell the system what it is seeing, limiting its autonomy.

What’s next?

Computer Vision seems to be the direction that Project Oxford is moving towards at present. They want to teach their programs to be able to recognize images in photographs and video and analyze what the main subject of the image could be. A speech recognition program that will eclipse those offered by Apple and their other competitors is also a priority.

IBM Watson

What is it?

Originally designed to defeat two grandmasters on the TV quiz show Jeopardy, Watson has since gone on to become the AI supercomputers by which all others are judged. The system can process over 500 gigabytes, or the equivalent of 1 million books, every second.

What can it do?

Watson’s original purpose was to beat two previous champions on the Jeopardy quiz show, which it was able to do convincingly after 5 years of design and testing. Now, IBM is partnered with over 300 firms from all fields, including Twitter, Wellpoint (Medical Insurance) and Chatterbox (Children’s Technology), to use Watson’s Natural Language Processing capabilities for their own ends. These partnerships continue to grow, particularly in the field of medical research and diagnosis.

What can’t it do?

For all of its power and memory, Watson’s capabilities are restricted mainly to Natural Language Processing, in other words, written and spoken text. This is because that is the format of the material for which Watson was originally produced. Therefore, Watson would not be able to analyze video at present, although there are plans for more cognitive functions in the not too distant future.

What’s next?

IBM has created a business specifically for Watson, and will be developing three new areas: big data visualisation, analyzing insights and R&D in the pharmaceutical industry. There are big steps being made in the treatments of cancers after Watson was paired with Memorial Sloan-Kettering Cancer Centre. There are also developments being made to help children, with Watson’s NLP capabilities being used in conjunction with Chatterbot toys to help interact with young children learning to talk.

Google DeepMind

What is it?

British tech company DeepMind Technologies, who had started to combine machine learning and the pursuit of neuroscience was bought out by Google in 2014. Having been renamed Google DeepMind, they are now tasked with building the best general purpose learning algorithms in the industry.

What can it do?

Unlike IBM’s Watson, which had a pre-determined purpose, DeepMind is more open-ended, and is also comparatively simpler to use than its competitors because of its deep learning capabilities. An example of this is that the system was given several vintage video games to play, with no prior input. The early results were obviously poor, but the system eventually managed to beat the games convincingly, having learnt from its mistakes.

What can’t it do?

The only disadvantage with deep learning capabilities and not having developers teaching the software what it should be looking for is that the whole process can be a great deal slower, and there are more likely to be erroneous results. However, in the future, the role of the developer in these systems will become smaller and smaller and more systems will look to adopt deep learning.

What’s next?

This is the million-dollar question. Since their success with the vintage video games, Deep Mind have almost gone into hiding, and appear to be doing a great deal of highly secret research, having hired a top Microsoft ecological modeller in November, although this gives little indication as to what their next step could be. We wait with bated breath…

Baidu Minwa

What is it?

Produced by Baidu, Google’s Chinese equivalent, Minwa is their landmark project, and mirrors the IBM Watson model, with over 72 processors and 144 graphics processors. Its image recognition capabilities are among some of the best in the world of artificial intelligence.

What can it do?

Much like Watson, Minwa is an image recognition engine, and a very powerful one too, with a 36-server node set-up, 6.9 terabytes of host memory and a 0.9 petaflop peak performance. Andrew Ng, one of the brightest minds in modern machine learning, and who once taught a computer to recognize cats, was directly involved in the design and function of Minwa.

What can’t it do?

Not unlike Watson, Minwa’s Natural Language Processing capabilities are some of the most impressive in the world, but the whole project was shrouded in disrepute after the most recent Image Classification Challenge, in which Minwa posted a 4.58% error rate, better than its competitors from Google and Microsoft, and better than the average human rate of 5%. However, the company revealed it had fallen foul of the competition rules, therefore nullifying their score.

What’s next?

It has been reported that Baidu’s most recent investment has come in the form of deep learning capabilities, which they are looking to use for Phoenix Nest, their ad-bidding platform. Although the company always keep a fairly tight lid on future projects, one could safely assume that the deep learning technology could soon find its way into Minwa’s architecture, removing the need for many of the developers inputting code and deleting abnormalities. This would also create the opportunity for Minwa to be involved in a more philanthropic role as Watson has started to be used. 

Comments

comments powered byDisqus
Key in hand small

Read next:

Unlocking The Potential Of Machine Learning

i