An excellent talk from DARPA on the three waves of Artificial Intelligence (AI) – The Contextual Adaptation is the right direction to go: Explainable AI

The AI technology (AIT) can be categorised in three main waves which are “handcrafted knowledge”, “statistical learning”, and “contextual adaptation”.

The first wave of AIT is the “handcrafted knowledge” which is all about the logical reasoning It can be implemented by taking the knowledge about a particular domain and charactrizing it in set of the rules that could be fit in the computer. The computer could then study the implication of those rules. An example for the first wave of AIT is the computer games like chess in which the output is clearly defined on the system based for each input based on the logical rules implemented by developer. The first wave of AI technology enables reasoning over narrowly defined problems however there is no learning and preceving capabilities while it is poor in handleing of the uncertainties.

The second wave of AIT is the “statistical learning” which is very good in preceving the surrounding world, for example in separating one face or voice from another. The are also very good in learning from datasets (vast amount of data for learning a new task) however they are not that good at resoning. So the second wave of AIT is powerful in classification and prediction tasks when provided with the context however there is no capability in undrestaning the context and they have got minimal reasoning ability too. So the detection/categorization of the context without undrestanding it is the main limitation which encourage us to move byond the second wave.

The third wave of AIT is about the “contextual adaptation”. In this wave the system itself will build over time the underlying explanatory models for classes and real-world phenomena. Just to comapre the second wave with the third wave just to comapre the performance of the second wave and the third wave in a image classification task. I give you an example. The second wave thinks an image is for example a “Cat” because it make all the calculation and the result for the “Cat” come the highest,  which is not that satisfactory. The third way system however thinks the animal shown in the picture a “Cat” because it has ears, it has paws, and it has pur and other features. So the the third wave of AIT can underestand the facts and features and not simply detect/chategorise but understands the context. So if we ask the system to tell us about the reasoning behinde this decision, the system  could provide the reasons in the way understadable to us. So it seems such a system constructs models based on the humanlike facts. So the thrid wave system of AIT will be build around the contextual models where the system learn how to construct over the time the models model precives. And It then preceives the world in terms of that model and it will be then able to use that model to reason to be able to make decision about things as shown in the following figure.

To watch the full version of the DARPA presentation presented by John Launchbury follow this video

Reference – https://www.darpa.mil/attachments/AIFull.pdf