AI3SD Video: Interpreting opacity: understanding gaps in our explanations of artificial neural networks

We know everything that goes on within artificial neural networks. We tend to know of all the data such systems have been trained on. And designers will be aware of the various design decisions, training algorithms and techniques that went into their construction, too. At the same time, leading AI designers tell us that their systems are in some sense uninterpretable, inexplicable or opaque. That’s puzzling. Drawing on discussions in the philosophy of neuroscience and science more generally, I will make use of this puzzle to try to advance our understanding of what explanations we lack with respect to ANNS; hence the nature and scope of explanation. The puzzle helps us to distinguish different phenomena in need of explanation, and some limits to the mechanistic explanatory strategies so often helpfully employed in the cognitive neurosciences