AI and machine learning continue to make headline news. And behind many of the sensationalist headlines are the consequences, or risks, of bad training data. Based in Dacorum, in the UK, Mindtech Global is the AI that can power AI and believes their products, Chameleon and the recently launched Dolphin, can provide not only the synthetic data, but also the analysis to ensure models are trained effectively.
Chris Longstaff, VP Product Management at Mindtech recently joined TFN to talk about Mindtech’s offer and their growth to meet the demands of the rapidly expanding AI sector.
Learning from the difficulties of AI
Mindtech was founded in 2018 by CEO Steve Harris and, at the time, was planning to offer a more straightforward AI product. “The idea was that we would create ML models and applications in the computer vision space,” says Longstaff. However, they soon came across obstacles in their path. “When we asked clients for training data, it was either ‘here’s a lot of data, but it’s unlabelled, and we don’t know whether we have rights to it,’ or ‘if you want data, you have to get that yourselves.”
Realising that addressing these issues while also meeting customers expectations would not be sustainable, they started looking for other solutions. The answer came from Peter McGuiness, now Mindtech’s VP of engineering, who suggested synthetic data, resulting in Mindtech’s pivot. The result was their first product, Chameleon.
Using synthetic data to improve real-life performance
Given the ubiquity of cameras in our daily lives, it might seem odd that there’s even a need for training data. Indeed, for those outside the sector, using generated data might seem a second-best option. However, Longstaff explains that using their synthetic data can lead to better results.
Longstaff uses the simple example of a retail store that wants to follow Amazon’s model of checkout-free shopping. “There are masses of cameras in those sorts of stores,” says Longstaff, “but a key problem is that packaging is constantly changing, you might have the same tin of beans, but some weeks it might have 10% extra free, some weeks it might have an offer price on it.” The result is that the system regularly needs retraining.
However, the problem is further compounded by shoppers. “You’ve got hundreds of cameras, but you don’t know quite where a person will stand, or what angle they’ll pick up the tin from.” But even that isn’t the end of it, says Longstaff, “if you train in the real world, you’ve got people, and you either need to get clearances from them, or hire actors.”
The result is that a cheap promotion on a tin of beans becomes a costly training exercise for the retailer, which, repeated across the product range, rapidly becomes impossible and unaffordable. Chameleon solves the problem by rapidly generating a model of the product that can be analysed in hundreds of positions and situations simultaneously.
Identifying the missing needles in millions of pictures of hay
However, despite Chameleon’s rapid modelling, machine learning can still be weakened by gaps in the data, and that’s where Mindtech’s second product, Dolphin, comes to the rescue.
“Talking to different customers, mostly they’d have a partially working system. They’d have real-world data, but not enough,” Longstaff explained. But identifying the missing areas could be problematic. Dolphin addressed that by analysing the existing data across a range of criteria to identify what extra data is needed.
The process involves three stages, Longstaff told us, “First, it looks at the content of the images and analyses that. Second, it looks at the appearance of the image, things like how bright it is and the colour space in it. Finally, we run machine learning models on it.” The results allow them to identify issues with the models, like false positives, or areas where more training is needed. And, of course, Chameleon can be used to provide that additional training data.
This combination of products is what Longstaff says sets Mindtech apart from potential competitors, “we’ve those three elements of analysis, it’s very rare to find anyone doing that detailed network analysis.” The ability to create synthetic data sets them further apart. “We can look at the data analysis and use that to set the parameters for Chameleon to generate missing data.”
What about funding?
Mindtech has yet to go through a formal funding round. Growing to have a presence in the US and Japan, as well as the UK, from its initial investments from names like Mercia Investments and the Northern Powerhouse Investment Fund. Currently in a pre-Series A phase, they are in discussions with existing investors, but opportunities exist for new investors too. But for Longstaff, the launch of Dolphin is the company’s current focus.
“We only launched Dolphin a few weeks ago, and already we’re seeing a lot of interest,” he told us. “The platform’s definitely staring to make people think about different uses which we hadn’t thought of, so we’re very excited about where Dolphin’s going.”
