Skip to main content

What is AI?

Machine learning

An algorithm that iteratively improves its performance. This happens by ‘training’ it to produce a desired output. The ‘training’ is provided by giving the algorithm large data sets and feeding back whether it has processed each data point correctly or incorrectly.

Deep learning

Systems that take an input and produce an output, which is then used as the input for the next layer of processing. Deep learning has produced ‘intelligent’ behaviours such as image classification and text recognition that were previously the domain of humans.

Artificial neural networks

Systems that emulate the workings of the brain. Code that can be fed an input and produce an output is linked to multiple other ‘nodes’. Each node is capable of both storing and processing data, meaning that every link in the network can process information simultaneously, vastly increasing computational power.

Search algorithms

Search algorithms retrieve information stored in a particular search space that meet set criteria. They are used in AI to find the path of least resistance from a start state to a defined end state, using particular paths. Search algorithms are used in tasks such as game playing.

Natural language processing

At a high level, the ability of computers to understand human speech and react in response. Natural language processing is the foundation of the Turing test of whether a computer is exhibiting intelligent behaviour.

The big challenge for our survival lies in the extent to which we can take control of data and AI. I personally feel a strong sense of crisis.

Kenichiro Yoshida, Chief Executive, Sony

Current and emerging use cases

Current and emerging use cases

Current and emerging use cases

Everyday AI

  • Search engines (Google)
  • Purchase suggestions (Amazon)
  • Email reply suggestions (Gmail)
  • Smart homes (Nest)
  • Intelligent assistants (Cortana, Siri, Alexa)
  • Facial recognition (Facebook, Apple, Alibaba)
  • Content moderation (YouTube, iQiyi)
  • Credit scoring (ZestFinance)

Industry, specialised and ‘edge’ cases

  • Diagnosing and treating diseases from images, device data and even brain waves (IBM Watson)
  • Reviewing legal documents to spot risks, eg uncommon or missing clauses (Kira)
  • Robotic logistics (Amazon, Alibaba)
  • Autonomous transport (Tesla, Waymo, Uber)
  • Care-bots and humanoid companions (Palro)
  • Self-learning game playing (AlphaGo and AlphaGo Zero)
  • Crime prediction (PredPol)
  • Debating (IBM Project Debater)
  • Creating a portrait like a renowned painter by analysing data points from more than 300 authenticated paintings, including lighting, angles, proportions and choices of subject (The Next Rembrandt)
AI by numbers

4.3x

0
.3x

growth in enrolment in AI courses at Stanford University between 2010 and 2017, the fastest in the US.

16x

0
x

growth in enrolment in AI and machine learning courses at Tsinghua University, China, between 2010 and 2017.

2.1x

0
.1x

growth in number of venture-backed US AI startups between January 2015 and January 2018. Growth in US startups as a whole over the same period was 1.3x.

5x

0
x

growth of AI patents in South Korea and Taiwan between 2004 and 2014, the fastest rate of increase in the world.

30%

0
%

of media stories about AI were positive in July 2016, up from 12% in January of the same year. Positive media sentiment has hovered around 30% ever since.

1000

0

the approximate number of games of Go, a complex Chinese board game, a human can play in a year. Google DeepMind’s AlphaGo – the first computer program to beat a world champion – can play many millions of games in a single day.

100-0

0
-0

AlphaGo was ‘trained’ using thousands of previously played games of Go. A later iteration, AlphaGo Zero, learned simply by playing itself, starting at random. It has since beaten AlphaGo by 100 games to 0.

.
Most US machine learning patent applications 2017
Top US autonomous vehicle patent applications 2017
.