API World + AI DevWorld

In recent years, a perfect storm of big datasets and advances in both the theory and practice of neural networks paved the way for many revolutionary applications and a wide-spread optimism about ‚Artificial Intelligence and machine learning. In particular, deep learning achieved tremendous adoption thanks to outstanding performances in a variety of tasks, especially in single prediction‚ problems: classifying images, predicting new values in highly non-linear series, etc. The not-so-hidden truth, however, is that behind any successful ML enterprise big data project, there are a lot of small data challenges that need to be solved: data need to be cleaned and structured in a context where supervised learning is next to impossible (organizations do not have thousands of examples to work with and solutions often involve a chain of inferences). In many cases, we cannot ‚ML-our-way‚ out of ML pipeline problems - so we are often stuck with humans, which unfortunately are costly and slow. The good news is that we all intimately know physical systems that are fairly efficient under that constraint, i.e. humans. We learn concepts with very few examples and then use simpler concepts to acquire more complex ones. We shall present a series of use cases from our experience working with large enterprises on how a different approach to learning and inference can bring superior results with less data, drastically reducing the time to ROI. In particular, we will present cognitively-inspired examples on lexical acquisition and program generation through probabilistic programming, showing how motivated priors can help navigate the space of hypothesis efficiently and produce a virtuous human-machine communication loop.