Why Link to External Academic Work?

In most engineering and science disciplines, the work contained in academic papers is rather far from being "ready for prime time", i.e. ready for application into commercial products. Having been through the implementation cycle of many new technologies we are very familiar with this lab-to-production gap.

 

However, things are a bit different in the new world of deep learning application & research. First, most research into new algorithms is completely open  - and due to the ease of replicating work in open source tools - most promising algorithms are heavily vetted in follow-up research by both the research community and those seeking to use these algorithms in commercial applications. Hence, in the deep learning revolution we find ourselves in the rather unprecedented situation of having a vast array of valuable new methods available as open and well studied starting points for building solutions.

To give an example in the field of medicine: The deep learning revolution in computer vision started in 2012, culminating in two open deep learning image classification network solutions of astonishing performance, called "ResNet" and "GoogLeNet" (from Microsoft & Google, respectively), and published in 2015. In 2016, these networks were used extensively by many competing research teams in the CAMELYON 2016 Challenge, whose goal was to accurately perform diagnostic assessments of lymph node metastases of breast cancer using deep learning to automatically read digital pathology slides. The resulting winning entry was by a commercial effort and close to the accuracy of human pathologists (and now has exceeded human performance). By using the determination of the pathologist + the algorithm a much higher accuracy (lower false positive and false negative) was achieved (see paper in link to right).

This remarkable transfer of deep learning's success from pure algorithm research to commercial solution has been repeated many time in the last 5 years, and so we view these academic references as a remarkably valuable resource and a starting point for finding vetted sources of the best and most replicable open work that has been reduced to an engineering tool in many cases. Papers referenced here that are not due to Paradigm Shift AI work are technology we can use as freely available starting points that we are capable of reproducing and in many cases have already vetted in our own prior work.