Major conferences in machine learning display thousands of research papers every year, and the numbers are steadily growing. Even though the problems those papers address have a common theme, their approaches may differ drastically. In broad terms, as researchers in the field, we tend to put those papers in one of the two categories: more scientific or more engineering-oriented. Attempting to clearly define what constitutes as science or engineering goes beyond the scope of the present proposal. Nevertheless, we discuss that the boundaries between the two disciplines have been blurring and argue that there is a need to marry important scientific and engineering practices to strengthen the outcome of our research.
After the 2019 edition of the workshop, the second iteration of Science and Engineering in Deep Learning aims at providing a venue where we discuss the values that define our community by deepening the communication channels among researchers working on seemingly contrasting challenges. Consequently, we hope to achieve stronger ties across various machine learning community subgroups, with the goal of fostering better integrated scientific and engineering practices that range from publication to production. We devote the first themed session of the workshop to discuss the following question:
What set of scientific and real-world values should we implement that would guide the theoretical and practical advances in deep learning?
The demands of our field change rapidly. We evaluate submitted research articles for their scientifically sound arguments; we ensure that the research is reproducible and robust under small perturbations. One might argue that a wave of scientific scrutiny is entering in the domain of engineering. In light of the tremendous social impact of deployed models, the necessity of such scrutiny couldn't be more evident. But this one-sided exchange between the two practices brings up the following question for the discourse: Shall we expect ethical science and engineering practices from scientists whose output is "pure machine learning and deep learning research"? We believe the answer is a solid yes!
This year, in the second themed session of the workshop, we plan to bring together experts to address the following question:
Why should machine learning researchers be concerned about the broader impact of their research?