Machine learning is not a synonym for AI
Mike Mullane wonders if Pepper the robot has an Austrian accent
There is a scene in James Cameron’s original Terminator movie where a flophouse manager pounds on the door of the cyborg’s room. Before choosing an appropriate response the Terminator scans a list of possible answers. At first glance, they seem quite different but Pepper the robot made me think of T-101.
For anyone who doesn’t know, Pepper is a semi-humanoid robot manufactured by the Franco-Japanese Softbank Robotics. Earlier this week, Pepper earned a minor footnote in history by appearing before a British Parliamentary select committee to testify about artificial intelligence (AI).
As the MIT Technology Review reports, Pepper’s testimony has upset a lot of people in the AI research community. Many have dismissed the event as a publicity stunt and worry that it will mislead the public about the current state of the technology.
Their concern is not misplaced. A respected British newspaper has already speculated on whether the Pepper the robot is pre-programmed to answer questions or if it relies on AI.
Here’s the thing: what we are calling AI cannot actually think for itself and only operates in highly controlled environments. One day Skynet may come, but for the time being machines are incapable of imitating human intelligence.
The experts differentiate between ‘strong AI’ and ‘weak AI’. Strong AI, sometimes called general AI, refers to a machine able to solve any problem requiring advanced cognitive abilities. It would be able to deal with new situations and solve problems it has never faced before.
Neither Pepper nor apparently T-101 possesses this kind of AI. They have weak AI, also known as ‘narrow AI’, which supports humans in solving problems in specific use cases.
AlphaGo, for instance, thrashed the human world champion of the board game Go, but would be useless at anything else. Virtual assistant like Alexa and Siri combine different weak AIs to create a kind of hybrid intelligence.
Virtual assistants can search the internet for basic information, schedule events and reminders and operate smart devices, among other things. This is not general intelligence.
‘Machine learning’ and ‘deep learning’ are two frequently used and often misunderstood terms.
In machine learning, the machine builds up the knowledge to complete specific actions based on training data covering multiple datasets. There are many examples of machine learning in our daily lives.
The performance of machine learning algorithms is directly related to the available information, which is referred to as ‘representation’. A representation consists of all the features that are available to a given machine and selecting the right representation is highly complex.
Representation learning is a field of machine learning that exploits raw data to automate the task of selecting the right representation. This is dependent on various environmental factors and far from simple.
For example, it can be difficult to distinguish colours in low light. Snowflakes, seagulls and shadows from trees are just a few of the things that can baffle self-driving vehicles.
Deep learning is a subcategory of representation learning that can transform features and elaborates dependencies based on inputs received. When a deep learning machine sees a picture, for example, it will map the pixels to the edges, to the corners and finally to the contours in order to identify an object.
In other words, deep learning is a kind of representation learning, which is a kind of machine learning, which in turn is a subcategory of AI. This should not detract from the very real achievements of the brilliant people who built Pepper. It is important, though, to understand what she is not.
In 2017, IEC and ISO became the first international standards development organizations (SDOs) to set up a committee to carry out standardization activities for artificial intelligence. Subcommittee (SC) 42 is part of the joint technical committee ISO/IEC JTC 1.
SC 42 is working with other JTC 1 subcommittees, such as those addressing the internet of things, IT security, and IT governance, as well as the IEC Systems Committee (SyC) for Smart Cities.
SC 42 has set up a working group on foundational standards to provide a framework and a common vocabulary. Several study groups have been set up to examine the computational approaches and characteristics of AI systems, trustworthiness, use cases and applications and big data.
Meanwhile, IEC TC 65 covers industrial-process measurement, control and automation.