Bay Vision - Spring 2018

During development, their team used Caffe framework a lot, which worked well to do the framework, but later on they moved to TensorFlow. They found it to be really strong, and could depend on the many libraries available. They also use their technology for in-phone deep learning inference. They have built the training methodology of first base product and are now moving onto the commercialization. They are also working on online learning which involves looking at how to use their technology to personalize shopping experiences. It remains a work in progress. Manindra explained more about the deep learning at scale: “ It really helps to know how to use cloud systems. Training on one machine is one thing. If you have to spread the training over multiple machines, you have to use millions of data points. Then the end result is actually, I would say, 50% or less of your deep learning knowledge and just 50% of all the full-stack engineering or cloud engineering. That’s really important for any team. That’s what we’re trying to do for deep learning. In our team, we are 50% deep learning and 50% full-stack. Data is definitely important, but where the industry is right now, there are basically two separate problems. One is collecting the data. That’s where the data scientist come in. The second is actually training them. Training them is a combination of some deep learning, but also a massive amount of full-stack engineering or cloud engineering . ” 17 GoFind Bay Vision