Technology advances have put an end to the 'AI winter'

Large data volumes, higher compute power and real-world challenges mean that AI investment will not slow down again, says Auto Traders' David Hoyle

Artificial intelligence and machine learning suffer from "a lot of hype" and are likely heading for failed expectations and disillusionment, David Hoyle - lead data scientist and head of data science at Auto Trader UK - warned attendees of Computing's AI & ML Live!

We're currently in an ‘AI spring' - a period of renewed investment and interest - largely driven by advances in deep learning, but does that mean we are headed for an AI winter? We have been through at least two in the last 60 years, but Hoyle argues that the situation today is not the same. The combination of large data volumes and higher compute power using GPUs has changed the AI landscape.

"One of the things that is different this time is advances in deep learning neural networks and the training of those networks," he said.

Deep convolutional neural networks (CNN) are a particularly popular type of DNN, mostly used for image classification; they break an image up into many small parts and, through a number of layers, construct features to learn what it is about the image that leads to the thing that it shows.

Because the latter layers of a CNN are "essentially constructing features that are generic to any image classification," pre-trained networks can be reused on different tasks using a very small amount of additional training data: examples are Google's Inception 3 and Inception 4 networks, which Auto Trader is experimenting with to identify car images.

It's not just Auto Trader; many companies are beginning to take an interest in AI and machine learning, with a "frenzy" in hiring experts ongoing; ‘data scientist' is the most clicked-on job on recruitment website Indeed by a factor of five.

However, wherever there is hype, there are challenges. "Data might be the new oil, but just because we have lots of oil doesn't mean we have lots of money," said Hoyle. A return on investment is notoriously difficult to achieve.

That difficulty, even amongst companies that have invested heavily into AI and ML, can lead to unhappy data scientists - especially if they were hired to do AI and end up working on something else (like machine learning, or linear regression).

Hoyle expects AI and ML to become cheaper and more commoditised in the near future, which will mean a change in employment: data scientists will "move to tackle new challenges, like genuine AI," giving space on the front lines to data engineers.

Hoyle called ‘genuine AI' "machines that can mimic humans - not machines taking over the world." That should be a relief for the technophobes.