If there's one thing people know how to do, and that's guess what caused something else to happen. Usually these guesses are good, especially when making a visual observation of something in the physical world. AI continues to wrestle with such inference of causality, and fundamental challenges must be overcome before we can have "intuitive" machine learning.
The cultural perception of AI is often suspect because of the described challenges in knowing why a deep neural network makes its predictions. So, researchers try to crack open this "black box" after a network is trained to correlate results with inputs. But, what if the goal of explainability could be designed into the network's architecture -- before the model is trained and without reducing its predictive power? Maybe the box could stay open from the beginning.
Software developers and cyber security experts have long fought the good fight against vulnerabilities in code to defend against hackers. A new, subtle approach to maliciously targeting machine learning models has been a recent hot topic in research, but its statistical nature makes it difficult to find and patch these so-called adversarial attacks. Such threats in the real-world are becoming imminent as the adoption of machine learning spreads, and a systematic defense must be implemented.
Machine learning algorithms are notoriously known for needing data, a lot of data -- the more data the better. But, much research has gone into developing new methods that need fewer examples to train a model, such as "few-shot" or "one-shot" learning that require only a handful or a few as one example for effective learning. Now, this lower boundary on training examples is being taken to the next extreme.