Cliff Bohm - Understanding AI from the perspective of information flow and information fragmentation

BEACON Seminar Feb 12, 2021 Clifford Bohm (MSU) Understanding AI from the perspective of information flow and information fragmentation While digital brains (i.e. structures capable of AI) continue to improve, they also are becoming more complex. This presents two issues. The first issue is a matter of trust. As brains become more complex, we lose the ability to investigate their inner workings. The second, and perhaps more interesting issue, deals with being able to continue the trend of improvements. As AI becomes more inscrutable, it becomes harder to know why different methods actually work. Fragmentation is a new measure, related to representation (R) which provides detailed insights into how data is stored in digital brains. The measure considers salient features of an environment and asks for the smallest subset of a brain's memory that contains that information. The calculations to determine fragmentation can also be used to determine how inputs and memory at one point in time predict outputs and memory and another point in time. This can be displayed as networks of data flow and these networks can be subjected to objective measurements - allowing for direct comparisons between different computational substrates (i.e. RNN vs Markov Brains...) and different training methods (i.e. backpropagation vs evolution).