Building Robust Machine Learning Models

Description
Sometimes, modern machine learning libraries can make building models look deceptively easy. Emphasis on tools and techniques like Python, R, and SparkML and deep learning seem to be common. However the speaker here argues that relying on these tools and techniques whilst ignoring the basics is the wrong approach in model building.

Machine learning in the real world requires discipline, hard work, and rigor. To develop robust models, it is essential that due diligence is carried out during the data acquisition phase and an obsession with data quality is required.

Choice of evaluation metrics, feature engineering, and a solid comprehension of the model bias/variance trade-off generally holds more importance than tool choice. Machine learning engineers with a lot of experience spend the majority of their time dealing with issues relating to data, parameter tuning, and model evaluation, whilst actually only spending a little bit of their time in building actual models. This is referred to as the 80/20 rule.

Unlike many machine learning talks nowadays, this presentation does not discuss deep learning. The hype around deep learning is being ignored to strictly focus on the main principals of building robust machine learning models.

Length: 1 hour 25 minutes