Bay Vision - Spring 2018

Hristo tells us that in terms of neural networks, they have certain stages in their pipeline that do matching and various specialized things have worked well, but the amount of training data needed for general classification can sometimes be an issue. When companies run promotions on particular products - like they have a special new year packaging for a product and that only runs for a few weeks and then goes away - recognising those can be difficult. He explains: “ It’s been more challenging to deal with that using these new approaches. There’s a lot of work that’s happening so I’m sure the fields will improve to where that’s reasonably good and practical, but we’ve had to actually revert to classical things, key points and that sort of stuff, to find matches and be reasonably accurate. ” In terms of their successes with computer vision techniques , Hristo remembers early on when they were trying out different approaches: “ One of the early implementations was we did classification in the oven. The idea was you put something in, we can detect what it is, and we can automate this process of detecting and making suggestions for you and just make it a very smooth user experience. That was one of the early applications. ” More recently, they’ve been doing more much more on the shopping storage part of the cycle. That’s early on in a user’s decision- making process and they can learn more about the user and provide value in terms of replenishment e.g. Would you like to buy this again? You’ve used this for this recipe, are you running low? It’s been this long since we’ve seen this particular box of milk in your fridge, should you do something with it? In conclusion, Hristo looks to the future: “ Our objective is really to reach scale , so for that the lesson has been just building systems that are practical. Carrying things out to production where you can support real-life data set sizes and make it effective, for example, running on AWS. Those are the kind of things that we’ve been really focused on .” 15 Innit Bay Vision “Our objective is really to reach scale, so for that the lesson has been just building systems that are practical”