After working in the Software and Semiconductor industries for more than twenty years, I and the early founders of Mindtech (who I’d known for a long time) got curious. Would it be possible, we wondered, to develop and deliver AI applications as a “Neural Network platform” on a chip. Fascinating though that was, we hit an unexpected roadblock. We were working with three lead customers and to train the networks, they each shared a huge amount of real-world visual data. But what we discovered when we looked at the data astounded me. Firstly, the data across all three customers was very, very poor. Most of it was completely unusable. Either it was not the right data or it was not in any way privacy compliant, with no data provenance. A big red flag. Secondly, the small amount of data that was usable took an inordinate amount of time to properly annotate because the process depended on humans. Real humans, painstakingly picking out pixel after pixel, hour after hour, day after day. It just wasn't scalable. And finally, when we looked at acquiring more data from bigger platforms, we got a hard no. Those platforms, if they had good data, knew the value and there was no way they were going to share it. In that moment, we realised that before we can take the next big leap in AI, we first needed to fix the issue of providing training data for AI systems. Real-world visual data is a scarce and precious resource. It has an important role to play in training visual AI systems, but we need an additional training data source that’s unlimited, unbiased, precisely annotated and above all privacy compliant. The idea behind Mindtech's Chameleon Platform was born. In 2019, we announced our end-to-end platform where companies can create synthetic 3D worlds to train their visual AI systems. The result is a real world where computers better understand the way humans interact with each other and the world around them.
