AI is reinventing the way we invent
www.nesta.org.uk/feature/innovation-squared/ai-reinventing-way-we-invent/
www.nesta.org.uk/feature/innovation-squared/ai-reinventing-way-we-invent/
Machines are rapidly getting smarter and can now accomplish feats that until recently could only be achieved by people – such as diagnosing certain diseases or driving cars. But what if, in the long run, the greatest changes from AI do not come from its application in day-to-day life but through altering the way in which innovation is managed and supported?
As described in a recent US National Bureau of Economic Research (NBER) paper, AI may serve '...as a new general-purpose 'method of invention' that can reshape the nature of the innovation process and the organization of R&D.'
Just as the greatest impact of optical lenses was opening new worlds with microscopes and telescopes – rather than helping us see through spectacles – so AI could change the world most by reinventing the way we invent.
One field where AI is already being put to work is that of new drug discovery and development. Here, as with other areas of research, one of the major challenges facing scientists is the sheer number of possibilities. Take drug candidates for example: the total number of small molecules that might serve as drugs is thought to be greater than 10 to the power of 60, more than the total number of atoms in the solar system. While these quantities might boggle the human mind, AI trained on data about existing chemicals and biological targets is well placed to explore this sort of problem.
At the same time, the conventional drug development pipeline is looking increasingly shaky. Estimates suggest that more than 85 per cent of drug development programmes do not lead to product approval. What’s more, it now costs over £1 billion to develop a new drug and takes, on average, more than 12 years. This has led many to question the current model of biomedical innovation.
Just as the greatest impact of optical lenses was opening new worlds with microscopes and telescopes – rather than helping us see through spectacles – so AI could change the world most by reinventing the way we invent.
Part of the problem may be that, even as mankind’s collective knowledge increases, the ability of individual scientists to connect relevant pieces has stayed relatively static. For example, according to Jackie Hunter of Benevolent AI, a new life science paper is uploaded every 30 seconds; that's over 30 million papers a year. Yet the typical researcher can only read between 200 and 400 in that time, let alone digest all the data from genomics or imaging. In short, there has been a recent explosion in biomedical data but our ability to process that information hasn’t kept up.
Benevolent AI uses AI tools to search the corpus of existing scientific information for previously unknown biological relationships to find new drug discovery and development opportunities. Its technologies shortlisted a handful of interesting ideas about how to address amyloid lateral sclerosis, also known as motor neurone disease – a form of which affected the late scientist Stephen Hawking. These were then tested at the Sheffield Institute for Translational Research. One of the ideas looked more promising than current treatments and is now being investigated further. Another idea had in fact already been proposed by the scientists at Sheffield, following two years of research – Benevolent AI had identified the same candidate molecule in two weeks.
Another area where AI is being applied to the process of innovation is material science. Innovative materials offer the prospect of better technologies in a whole range of areas, from better batteries to more efficient solar cells, to catalysts that make fuels from artificial photosynthesis. However, the traditional process of materials development is often slow: Tonio Buonassisi, a mechanical engineer from MIT, says: 'It takes an average of 15 to 20 years to come up with a new material'.
Like drug discovery, identifying suitable materials to solve real world problems can be difficult. Not only are mind-boggling numbers of new chemical compounds possible, but these possibilities are made increasingly complex by phenomena like polymorphism (where a solid can have more than one crystal structure), by the combination of different materials, and by the introduction of nano-scale features. This can give rise to wholly new, emergent properties.
To help explore this vast ‘parameter space’, researchers in academia and industry are using machine learning to scour data sources, gain novel insights and build new predictive models. Combined, this is helping to identify new materials such as more corrosion-resistant metals, improved membranes for fuel cells and compounds that absorb pollution.[1]
What makes AI so interesting as an innovation method is not so much its impact on particular domains of research, such as pharmaceuticals or green technologies, but how widely it can be applied. Artificial intelligence, and deep learning in particular, can be thought of as both the 'invention of a method of invention' and as a technology with general purpose use across many types of innovation.
For example, functional Magnetic Resonance Imaging (fMRi) has transformed our understanding of the brain both through the insights it has generated and as a new tool for this sort of research. Its development can therefore be characterised as 'the invention of a method of invention'. But despite its value in brain research, fMRI has few uses in innovation beyond this field, and so does not have general purpose. In contrast, certain forms of AI can be used as a tool for innovation that, as AI pioneer Geoffrey Hinton and colleagues described, are ‘applicable to many domains of science, business, and government’.[2]
What makes AI so interesting as an innovation method is not so much its impact on particular domains of research, such as pharmaceuticals or green technologies, but how widely it can be applied.
Is there a danger that AI may threaten the jobs of innovators in the same way that some believe it does certain less-skilled professions? Leaders in the field of AI and innovation think not: Jackie Hunter has described AI as something that would augment researchers rather than replace them, allowing them to focus, save time and potentially make less biased decisions. Others are even more enthusiastic about the opportunities for AI and people to work together: as Neil Lawrence, an AI expert from the University of Sheffield, says: 'Computer empowered humans operating in synergy with machines? Wow!'
That is not to say that there won’t be bumps ahead, however. Management experts Iain Cockburn, Rebecca Henderson and Scott Stern have flagged one of the potential challenges of AI-driven innovation: a race for data. Many recent advances in AI require enormous datasets, and so many of those involved have a strong incentive to gather and hoard data. First movers could gain major long-term advantages through control of data, with new entrants unable to challenge incumbents. In these circumstances, without action to encourage sharing, the benefits of AI as an innovation tool may not be fully realised.
Another, and potentially more profound, challenge for the use of AI in innovation is whether its results can be reproduced. At a recent meeting of the American Association for the Advancement of Science, Dr Genevera Allen from Rice University in Houston described how the increased use of machine learning techniques was contributing to a 'crisis in science'.
Software used to analyse data that has already been collected is sometimes producing answers that are wrong because the patterns identified do not exist in the real world – and this problem is exacerbated by the relative opacity of some machine learning systems. Allen said: ‘People have applied machine learning to genomic data from clinical cohorts to find groups, or clusters, of patients with similar genomic profiles… but there are cases where discoveries aren't reproducible; the clusters discovered in one study are completely different than the clusters found in another’.
To counter this problem, a number of programmes such as the Defense Advanced Research Projects Agency’s Explainable AI (XAI) and Competency-Aware Machine Learning (CAML) are developing tools to get machines to explain how they reached their conclusions.
Is there a danger that AI may threaten the jobs of innovators in the same way that some believe it does certain less-skilled professions?
AI’s universal applicability could make it a key solution to a potential productivity problem in innovation. Evidence from a wide range of industries, products and services shows that while total research effort is rising, research productivity is falling.
For example, a group of eminent economists point out that, ‘the number of researchers required today to achieve the famous doubling every two years of the density of computer chips is more than 18 times larger than the number required in the early 1970s.’ AI can potentially assist by searching for useful ideas among the vast number of possible research avenues generated by previous research, and by drawing researchers’ attention to patterns and connections which they might otherwise overlook.[3]
The use of AI for innovation is not new, but advancements in its forms, combined with new access to data, means there has been a step-change in the opportunities offered by this technology.[4]
The ability of AI to analyse vast quantities of data of different sets and suggest ideas beyond the human imagination certainly shows great promise as an innovation method. As an article for Government CIO Media and Research describes: 'It’s more than just crunching numbers, but also AI’s ability to follow unexpected threads.'
Despite huge potential and real progress, AI as an innovation tool has yet to experience an 'AlphaGo moment' – when a machine far surpasses the ability of people – but that may be just around the corner.
[1] Schleder, G.; Padilha, A.; Acosta, C. & Costa, M.; Fazzio, A. (2019). ‘From DFT to Machine Learning: recent approaches to Materials Science – a review’, Journal of Physics: Materials. doi://10.1088/2515-7639/ab084b.
[2] Yann LeCun Y, Yoshua Bengio Y and Geoffrey Hinton G (2015) Deep learning. Nature 521, 436–444.
[3] & [4] Agrawal A, Gans J and Goldfarb A (2018) The Economics of Artificial Intelligence. NBER, USA.
