Digital Catapult begins accelerator programme to help AI startups overcome major hurdles

Innovation centre Digital Catapult has launched a business accelerator programme to boost artificial intelligence (AI) startups.

The three-year programme, called the Machine Intelligence Garage, will provide AI and machine learning companies with access to computational equipment and expertise. 

Marko Balabanovic, chief technology officer at Digital Catapult, said the project aimed to help startups overcome current hurdles and build their presence.

“What we’re hoping to see is that [the startups] will get their products out more quickly and will create better machine learning tools and get to their customers more quickly,” he said. “They’ll basically get to the market quicker and be able scale up their company quicker.”

There are five startups in the first group, and a new set of startups will join the programme every six weeks. Each new cohort will receive “intense access to the expertise and the equipment” until the next group arrives, said Balabanovic.

All member companies will remain part of the programme for the full three years, with access to further mentoring and events. 

Balabanovic said there were four significant hurdles for AI startups, which are stopping them from growing. One of these is the cost of computational power. As the technology has become more advanced, more computational power is needed, but this comes at a cost that startups often cannot afford.

“You need a staggering amount of compute power to run this kind of stuff. You’re looking at these little startups having to spend tens of thousands of pounds on cloud compute to do one training run,” he said.

Digital Catapult surveyed 68 of around 400 AI startups in the UK, and found that 60% of respondents felt computation-constrained.

The second barrier Balabanovic listed is getting staff with the right skills, as startups often need to compete with larger enterprises.

The third problem is gathering suitable data to train the machine learning program. AI needs vast amounts of data to learn from, and startups can find it difficult to access data, again because smaller companies do not have the same resources as large organisations.

Balabanovic said the final challenge surrounds the regulatory and ethical issues of AI. Startups could inadvertently implement prejudice, he said, or face questions about how the system came to a particular conclusion.

“There’s the regulatory side and the ethical side. How do you make the system function well in the real world, both in the legal sense and the ethical sense?” he said.

“It’s quite challenging to know: Have you accidentally introduced some sort of bias? Can you explain how the decisions are made in your system? A lot of these systems are black boxes.”