editorial

More than machines

U
nderstanding human thought has for years have seen remarkable demonstrations machine intelligence (such as deep
centuries kept the brightest minds of deep learning performing tasks that learning systems) to specific areas
busy, with many of the key questions humans are typically skilled at: playing — and topics in other domains (for example,
remaining unanswered. This hasn’t and winning — strategic games, medical physics, biology and healthcare); and
stopped us from imagining machines diagnosis at the level of human experts3, lastly the study of the impact of machine
that imitate our thoughts and behaviour. speech recognition and language translation. intelligence in society, industry
Such visions of intelligent machines On the one hand this form of AI is and science.
have influenced how we contemplate considered ‘narrow’, because deep learning This first issue of Nature Machine
our humanity and have given us popular systems are typically trained for one Intelligence features pieces in each of these
narratives such as those pondering the specific task. However, scientists from a categories. In an Article, Etienne Burdet
coexistence of androids and humans, or wide range of fields are becoming aware and colleagues report on work of the first
dystopian tales where a super-intelligent that the impact of deep learning is far from category, on developing algorithms for
power has taken over the world. More ‘narrow’: a well-thought-out application adaptive human–robot collaboration,
pervasively, the pursuit of machines that could be transformative for a field, tackling and the Review by Kenneth Stanley et al.
mimic some of our intelligence and carry fundamental questions and enabling fast presents an overview of approaches that
out useful tasks has played a key role in progress — for example, in finding new combine deep learning with evolutionary
shaping the technology that underpins materials, identifying important signals algorithms. For the second theme, one
societal transformations, from the industrial in particle physics or astrophysical data, of the best-known applications of deep
revolution to the digital age. discovering links between genetics and learning is in medicine, and a Perspective
The development of intelligent machines phenotypes, and in classifying cellular by Edmon Begoli and co-workers examines
seems to be gathering pace, as we regularly tissues in microscopic images. the need for uncertainty quantification in
hear about an artificial intelligence (AI) With every advanced technology that this area. In a Comment, Stephen Cave and
revolution. However, contrary to reports, becomes available, the risk of misuse and Seán ÓhÉigeartaigh discuss the different
we are not witnessing some sort of unwanted societal implications needs to timescales of ethical concerns in AI, falling
technological singularity: the eye-catching be examined. These ethical issues are all in the third category.
advances from the past few years are largely the more relevant for AI given that it is We also start a series called Challenge
a continuation from decades of sustained a decentralized phenomenon, in which Accepted, which are reports on data
research and funding. For instance, self- substantial power has been handed to challenges and competitions in AI and
driving cars autonomously sharing the road those companies and institutes with natural robotics, to highlight the important
with other traffic were already demonstrated access to much data. Ethical concerns and sometimes surprising role these
in a test in 1994 in Paris1. That machine around the development and use of play in steering a field, giving young
learning would be of great industrial use was AI are taken increasingly seriously. Such researchers a chance to demonstrate
also well understood in the early 1990s2. concerns have led to widespread efforts to their skills and to crowdsource solutions
A telerobotic spacecraft landed on Mars develop laws, principles and standards to to outstanding practical questions in
and sent back pictures in 1997. promote the human-friendly use of AI and science or industry.
This is not to say that there hasn’t been to start research programmes studying its The pursuit of intelligent machines will
a shift in impact. There are several factors societal implications4–6. continue to inspire in many ways, providing
underlying this shift, such as the advance There is no doubt that the deeply us with insights into human intelligence
in computer processing power, but perhaps ingrained desire to understand and as well as stimulating technological and
most importantly the availability of huge mimic intelligence will continue to scientific innovation that could lead to
amounts of data since the late 1990s. inspire technological and scientific future societal transformations. Now is the
These developments were in turn due to innovations. Nature Machine Intelligence time to be part of the conversation. ❐
the spread of affordable mobile devices that will endeavour to bring different fields
collect data on a massive scale, the ability to together, forging new collaborations Published online: 7 January 2019
share data as well as the accelerating storage in AI, robotics, cognitive science and https://doi.org/10.1038/s42256-018-0014-z
capacity. As a result, deep learning, which machine learning, to further develop
can recognize patterns in complex datasets visions of intelligent machines that can References
1. Delcker, J. Politico https://go.nature.com/2QOfrxP (24 June 2018).
when there is sufficient data to train the be of inspiration and use for humanity. 2. Jordan, M. Medium https://go.nature.com/2Bedwss
system on, made its famous breakthroughs There are roughly three main themes that (19 April 2018).
around 2012. Deep learning is considered we will initially focus on: the engineering 3. De Fauw, J. et al. Nat. Med. 24, 1342–1350 (2018).
4. AI Now; https://ainowinstitute.org
as a form of AI, as it bears similarities to and study of algorithms and hardware to 5. Human-Centred AI; http://hai.stanford.edu
human-based learning. Indeed, the past few build intelligent machines; applications of 6. AI4People; http://www.eismd.eu/ai4people

Nature Machine Intelligence | VOL 1 | JANUARY 2019 | 1 | www.nature.com/natmachintell 1