Rajat Thomas’ academic journey meandered, but data was always at the core. From an electrical engineer working in image processing of data from Large Radio Telescopes to a PhD in Astrophysics focusing on the early universe, Rajat was immersed in Big Data. An interest in the human brain and neuroscience brought him to the Netherlands Institute for Neuroscience, and these days he’s diving into pure artificial intelligence and machine learning (ML).
He joined us to share insights on the cutting edge of Bayesian Deep Learning with our data specialists.
Bayesian Neural Networks are a method of incorporating prior knowledge into a machine learning algorithm to assign a probability to predictions. The relevance of Bayesian Neural Networks stems from a fundamental problem in ML. When we look at ML Algorithms, we often don’t know why they make errors. Algorithms work in high geometric and mathematical dimensions, so when they make a mistake it’s often impossible for humans to decipher what their reasoning was. For example, we could feed a blank image into an image-recognition ML algorithm and it might tell us it’s a panda, but we won’t know why it deduced that. By incorporating Bayesian learning into ML we can ask an algorithm to tell us when it’s unsure. It might say, ‘it’s a panda but I’m only 1% certain,’ and we would be better able to disregard the conclusions. When we’re talking about a single image the importance of establishing uncertainty may not be the first thing that comes to mind but often these ML algorithms are feeding conclusions to more complex algorithms or to upper level management who uses the feedback to make crucial business decisions. At that point understanding uncertainty becomes important.
Rajat’s talk focused on the concept of uncertainty. Specifically on the evidence term which is how we understand the world and the data that we have. Models are inherently reductionistic. If we want to model the probability that someone walks past our window, to achieve the most accurate answer we would want to know where every single person in the world is.
In reality we build a model based on a small chunk of information (information from a camera we’ve placed outside our business for the past few weeks). A model can give us its uncertainty, but there is also a level of uncertainty in the very approximation we create, the information we feed into a model. He pushed us to challenge ourselves and ask: How can we create a better approximation of the world? For Xomnia’s David, the talk allowed him to take an intellectual step backward, look at the technical solutions that we are applying, and think deeply about algorithms. “These talks allow us to break outside the barriers sometimes created by over-focusing on practical applications.
Rather than searching for the best tool in the toolkit we’re able to think about inventing new tools. That creativity in technical approaches is what keeps me at Xomnia. Xomnia’s Yu Ri found that the talk changed the way that he thinks about a current logistics problem he is working on. He’s using a Bayesian approximation (true Bayesian neural networks require a great deal of computing power and aren’t ideal for the rapid results our client wants) to better model uncertainty in a logistics application he is developing. His model currently pulls data such as volume, transit data, and weather to make live predictions on when and where the logistics operation will be delayed.
The Bayesian application is particularly important because when there is a high degree of uncertainty in the model, the client wants to trigger human intervention. Rajat got him thinking about what it really means for an algorithm to share its uncertainty. We mentioned earlier that a model is really just a simplified representation of reality. For him the takeaway is that for the Bayesian inferences to have relevance, his focus needs to be on calibration of the model so that it more accurately represents reality. An uncertainty value is only as valuable as the model that you create. Bayesian Neural networks are still in their infancy. There are only a few well-established methods for incorporating these principles into technical applications. Our hope is that Xomnia can extend the research, and find creative applications for cutting edge technology to give our clients a leg up. Rajat has returned to Xomnia for the second time to dive more into technical applications and how deep learning can help Bayesian methods.