Autor: DPP

Artificial Intelligence is For Real (1 / 3)

No, AI isn´t just a hype.

The Context

Artificial Intelligence (AI) reached the peak of its hype curve in the consumer technology market as early as January 2017, when almost everything at that year’s Consumer Electronics Show was labelled ‘intelligent’, even though very little was – or would ever need to be. There is a closer than ever link between the hype topics of the professional media industry and those of the consumer world, but the professional sector still tends to lag a year or two behind its consumer counterpart. And so it was that the hype of CES 2017 wasn’t replicated in the media industry until 2018 – when many left NAB and IBC mumbling that much was being claimed for AI, with scant evidence to support it. Hype almost invariably begets disillusionment; and neither are helpful to anyone attempting to plan a business or assess a market.

This tension is particularly unfortunate in the case of AI since – unlike most over-hyped technologies – it seems almost inherently to offer enormous business benefits. The opportunity to ‘do more, better, for less’ has been a mantra dangled in front of many company boards over the last decade or so. It is a promise rarely delivered on, but the successful application of AI could, in theory at least, change all that. After all, AI is all about tackling or initiating processes that are beyond the capabilities of people.

It seemed particularly timely, therefore, for the DPP to bring together a group of subject matter experts from its membership to make a real-world assessment of exactly how AI is being applied in the professional media industry today; what the opportunities are for the next stage in its application; and what challenges represent the greatest impediments to further adoption.

The output from those discussions is summarised in this report.

The Definition of Terms

The term AI is applied in a huge range of different ways. In bringing together a group of AI experts, it would be easy to initiate a somewhat theological, and entirely unproductive, debate about what constitutes ‘true’ AI.

It was important therefore that the experts at our DPP AT HOME event should agree and employ a common understanding of the term AI.

The DPP invited Lydia Gregory, from AI consultancy FeedForward AI, to offer a definitional framework for such a common understanding. The framework she put forward can be summarised as follows.

A broad definition

For the purposes of the DPP AT HOME session, precise definitions of the term artificial intelligence don’t really matter.

Humans have long been obsessed with how they might enable the rules of a computer to capture something of what it is to be human by performing certain human processes. It’s a concept that began with Ada Lovelace and Charles Babbage’s Analytical Engine in the 1840s. It developed with the concept of neural networks explored in the 1940s. And then the term ‘artificial intelligence’ was coined in 1956 when John McCarthy invited ten scientists to come together for a summer so that “an attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves.” McCarthy defined artificial intelligence as the science and engineering of making such intelligent machines.

McCarthy’s definition still works well. What is more interesting and useful for a session such as this, however, is the practical business application of AI. When talking about AI we are essentially talking about automation – whatever the business happens to be. And automation falls into two forms: rules based automation, and machine learning.

Rules-based automation

A rules-based approach to AI involves capturing something that is intrinsically contained within humans. There are two versions of this.

One approach is where humans perform repetitive processes, for which a set of reproducible steps could be defined. This is Robotic Process Automation (RPA). An example of a business process that would benefit well from RPA would be one in which data is entered several times into separate systems. If the way a person performs those processes is captured, then a system can subsequently perform them automatically.

Another rules-based approach is where humans have expert knowledge. An example would be the expertise required to play chess. This is knowledge captured in the human brain. If that person works with a programmer, their knowledge can be codified as rules. This approach is not very scalable, but it is still very valuable for many businesses.

Machine learning

Machine learning (ML) approaches enable the computer to learn from data, rather than only executing pre-defined rules. ML can also be subdivided into two types:

The first is where the algorithm is trained from a static data set from which patterns can be learnt. The machine is then able to apply this learning to other data, for classification, the automatic application of tags, the ability to spot patterns, the ability to predict additional data points in a series, and so on. This type of model is good for up-scaling the resolution of images, for example.

The second type of machine learning is one in which the machine learns not just once from a training data set, but also continually from feedback as it is used. This is often referred to as deep learning. An example would be a video recommendation engine that responds to user feedback. The engine may record, for example, that the user doesn’t ever select certain kinds of content, or notice that the user does respond well to particular content when promoted in certain ways. The engine would revise its personalised recommendations, and reconfigure the recommendations interface as a result.

Deep learning may include the use of neural networks – an integrated group of individual AI processing units, which mimics the network of neurons in the human brain. An example of the application of neural networks is voice search, using tools such as Alexa, Google Assistant, Siri, Cortana and so on.

An agreed AI framework

Significantly, our group of experts were happy to adopt Lydia Gregory’s framework, without debate. For the purposes of the session, the various types of AI described by Lydia were categorised, and labelled, as follows:

Type A: Rules-based
A1: Hardcoded rules
A2: Robotic Process Automation

Type B: Machine learning
B1: Machine learning (not informed by feedback)
B2: Deep learning (informed by feedback)

Read part 2 soon on DPP Film Tech App.

Related Articles: