A branch of machine learning is drawing inspiration from the theory of evolution to construct more powerful algorithms. The ‘evolutionary’ technique is helping to design computer software programmes that are able to learn new tasks more easily.
At the EmTech Digital event hosted by MIT Technology Review in San Francisco on 27-28 March, Pedro Domingos, a Professor at the Department of Computer Science & Engineering at the University of Washington and author of a book entitled ‘The Master Algorithm: How the quest for the ultimate learning machine will remake our world’, outlined the five main schools of thought in machine learning – the symbolists, the connectionists, the Bayesians, the analogisers and the evolutionaries. The evolutionary approach has its roots in biology and in the theories of Charles Darwin. In fact the Deep Learning techniques in use today draw overall inspiration from the way the brain functions, but the brain itself has of course been subject to evolution over time, Domingos underlined. “The idea of the evolutionaries is to draw inspiration from the way evolution works in order to apply the same logic to computer programmes as for living beings. Artificial Intelligence systems are generated using random genomes. We test them, we keep those that are best suited to the tasks they’re designed to accomplish, we use their genome for the next generation, and so on,” he told the audience
Zachary Chase Lipton, a researcher specialising in Artificial Intelligence (AI), pointed out that the idea of using evolutionary algorithms is an old one, which had gradually fallen into abeyance but is now coming back into fashion due to the development of deep learning techniques. “This method is in fact very close to a classic neural network. The difference lies in the way we align the parameters in order to increase programme efficiency. When we work with a traditional neural network, we make calculations so as to optimise our gains at each attempt. With an evolutionary strategy, we apply a large number of random mutations to the parameters and select the most effective ones,” he told the EmTech Digital audience. The history of ‘evolutionary’ AI goes back to the 1950s, to the work of John Kosa and John Holland. At that time, this approach was restricted by the limitations of the available hardware, but the technique now looks likely to reclaim centre stage, reckons Zachary Lipton.
AI taking takes the controls
Ilya Sutskever, Research Director at non-profit AI research company OpenAI, which was founded by Elon Musk and Sam Altman, also presented his work on evolutionary AI at the EmTech Digital event, pointing out that this technique is particularly effective when it comes to teaching computers to carry out complex tasks that require several steps in order to achieve a result. Sutskever and his colleagues have tested a number of approaches by training a software programme to play around fifty different games on the Atari 2600 home video game console, which came to the notice of people beyond the circle of nostalgic gamers in 2013. At the time, UK-based startup DeepMind was using ‘reinforcement learning’ – a branch of machine learning – to develop software capable of beating the top human gamers. Shortly afterwards, following its acquisition by Google, DeepMind developed the AlphaGo software, which became famous all over the world when it achieved victory in March 2016 over Lee Sedol, one of the best ‘Go’ players in the world.
Ilya Sutskever claims that by using the evolutionary approach he has now achieved results similar to those DeepMind achieved with reinforcement learning, but in less time and at lower cost. He told the audience that in one hour his team can design a programme that would require a full day’s training on DeepMind’s methods to achieve similar results. According to Sutskever, evolutionary techniques also enable the design of programmes which can switch easily from one specialism to another, thus opening up the potential for more widespread use of Artificial Intelligence to carry out any task, successfully managing different types of complex scenario. “If you manage to build a generalised learning algorithm and constantly improve it, you can then solve a huge number of problems,” he underlined, although acknowledging that this technique is still in its infancy.
The Atari console
More versatile AI
OpenAI is not however the first to explore the potential of this technique. The pioneer in this field is Frenchman Antoine Blondeau, who in 2008 co-founded Sentient Technologies, a company that is quietly working on developing broad evolutionary Artificial Intelligence systems. “The AlphaGo and Siri systems, and those used in financial sector software, all have one thing in common. They excel in solving a very specific problem, but can’t do anything else. Our aim is to build a system that can solve a large number of problems, that can actually understand the world better,” Blondeau revealed.
The range of potential applications of Sentient Technologies’ algorithms would appear to be limited only by our imagination. The company is working in the online commerce sector, helping to change the way we shop. Its algorithms allow people to link clothes culled from pictures in fashion magazines or outfits photographed in the street with offers from online shops. You can for instance select an item you like from a picture and receive a suggestion for similar items that can be bought online. So you would no longer have to browse through masses of different products for hours in the hope of finding that rare gem. “Shopping online is thus getting more like shopping in a store, it’s becoming iterative and interactive. And Artificial Intelligence is acting as the sales assistant,” Antoine Blondeau told the auditorium.
Meanwhile on the healthcare front, Sentient Technologies has developed a technique for predicting the appearance of septicaemia thirty minutes before a doctor is able to spot it. The quicker blood poisoning is identified upstream, the greater the chances that the patient will survive. The method still has to obtain the official approval of the medical authorities, but when it does it looks certain to help save many lives.
Blondeau’s company is also working on behalf of the agriculture sector, where “In partnership with the MIT Media Lab, we’ve developed a system of smart containers. Inside, Artificial Intelligence is used to control the lighting, humidity, temperature, plus also the nutrients given to the plants, thus supervising their growth in real time. This means we can increase effective yields without having to resort to GMOs,” explained Antoine Blondeau. “
In the field of cybersecurity, the evolutionary algorithms developed by Sentient Technologies enable security faults to be identified before they can be exploited by hackers. In manufacturing, they enable predictive maintenance, so that machine parts can be replaced before they fail. In the insurance business, they are helping to draw up customised cover policies in line with the client’s precise needs. Last but not least, under Sentient’s partnership with MIT, the company is working on predicting seismic shocks.
By contrast, Ilya Sutskever, did not reveal exactly in which fields OpenAI is planning to apply its evolutionary algorithms, saying only that they would help to raise the profile of the concept of Artificial Intelligence. “If we can build computing systems capable of learning to carry out complex tasks all over the world, then we’ll really be able to call them ‘smart’,” he stressed. Meanwhile Pedro Domingos pointed to a field likely to enjoy high public visibility, underlining the potential of such techniques in developing robots able to move around autonomously.