14 AI IN THE OPERATIONAL MANAGEMENT OF LARGE-SCALE IT SYSTEMS
intelligence. Researchers from Vicarious AI have developed a computer vision model which is 300-fold more
data efficient at resolving text-based CAPTCHAs than deep neural networks. The proposed model is capable of
generalising from a small nu mber of training examples in performing a range of visual cognition tasks31. The
proposed approach, incorporating inductive biases from systems neuroscience in the development of computer
vision systems, can lead to the development of generalisable machine-learning models operating with high data
efficiency. The authors suggest that building on this work and combining neural networks with structured
probabilistic models may eventually lead towards general purpose AI systems. In the near term, however, new
approaches to improving t he efficiency of AI systems could lead to the development of high quality on-device
AI applications, which may be of direct relevance to addressing the challenges of the JHA community , such as
highly precise biometric on-device recognition and identification.
3. AI applications relevant to eu-LISA:
opportunities and challenges
If you have followed the news over the last couple of years, you have no doubt seen at least a few headlines
about AI outperforming humans or the potential for AI to displace millions of humans from their jobs. While the
former sounds very promising and the latter rather alarming, the reality is likely somewhere in between. As
already described, existing AI performs extremely well, indeed exceeding human performance, but only at a
very narrowly defined task, e.g. playing Go or recognising patterns in text or visual media. Whenever the task
becomes more complex and requires contextual understanding, the machine will likely either fail at the task or
perform substantially below humans. It is therefore unlikely that the current approa ches to AI, predominantly
based on machine learning, will be able to replace humans entirely in the near future32. Where AI can indeed
have a significant effect is in augmenting human performance by either providing s upport with information
processing and providing inputs to decision making, or performing routine tasks and freeing up time for humans
to engage in activities that require human involvement.
Davenport and Ronanki33 offer a useful categorisation of AI technologies based on business capabilities:
Process automation: automation of physical or digital tasks primarily administrative bac k office
and financial activities using robotic process automation. These tools can be used, for example, in
processing requests received by an IT service desk.
Cognitive insight: detecting patterns in large data sets and across numerous data sets, and
interpreting their meaning. Cognitive insight systems are largely based on some kind of ma chine
learning algorithms and are used for predicting outcomes based on data.
Cognitive engagement: customer engagement using natural language processing and generation.
These systems often take the form of a chatbot or an automated voice-c ontrolled assistant (e.g.
Apples Siri Google Assistant or Amazons Alexa
In addition to t he above, computer vision is being increasingly used in the public sector. Comput er vision has
been successfully applied in autonomous driving, detection and diagnostics of cancer in healthcare, detection
and monitoring of oil spills in the oceans, as well as safety and security applications including border
management. In the table below we provide a brief overview of AI applications in the public sector (see Annex I
for a more detailed overview of possible applications).
31 George et al. (2017) A generative vision model that trains with high data efficiency and breaks te xt-based CAPTCHAs. Available online:
32 Some notable exceptions include autonomous driving and some areas of medical diagnostics whe re AI is already on a par with, or better than, humans.