When building Machine Learning products, developers often encounter challenges that can waste both time and money. Common pitfalls include complex model pipelines, which can lead to exponential increases in development time. Data tagging, typically a tedious job, can be reduced significantly when using techniques such as Transfer Learning. In many cases, solutions that show great promise in development or test cycles fail to materialize in production. Data Driven Development (DDD) can create a more focused development cycle, reducing unwanted surprises when the model is released.
In this presentation, Ohad Zadok will focus specifically on the Machine Learning challenges developers face. Ohad will provide methodologies that will enable the scalability and performance needed to succeed. The presentation will also highlight examples of Machine Learning failures and how seemingly promising solutions may never reach expected levels of success. The lecture is based on Ohad's experience in a Machine Learning focused startup, which grew from six employees to more than eighty employees during his tenure, and this Google research paper "Hidden Technical Debt in Machine Learning Systems".