Where Deep Learning Fails
Recent progress in deep learning has garnered considerable excitement throughout the data science community. Many now believe that to improve a model's performance, you just need to "throw more deep learning at it". In this talk, I'll explain why such thinking can be costly. I'll highlight some notable cases where deep learning has failed to improve performance compared to simpler approaches. In particular, I'll deep-dive into the problem of modelling structured data, which is typical across retail, marketing and more. I will try to provide some intuition about why deep learning struggles in these domains and suggest some alternative approaches to try first.
Adam is a Senior Data Scientist at dunnhumby, where he builds and deploys machine learning algorithms at scale for some of the world's largest retailers. He is also a part-time Experimental Psychology PhD student at UCL, funded by dunnhumby. His research aims to better understand customer purchase behaviour through a combination of data science, machine learning and cognitive modelling.