In this 45-minute talk at the University of Southern California, Dasgupta looks at the progress of supervised learning over the years and poses a question about how supervised learning algorithms and techniques can be adapted to interactive learning, in which a machine interacts with the human operator. He asserts that the success of such a method would depend upon various factors, particularly the following used by the machine: (1) the data clustering algorithm; (2) the query strategy, because all data cannot be presented to the operator for him to classify and label; (3) the protocol between machine and human operator; (4) the incorporation of feedback, particularly that on features of data; and (5) the coining of suitable cost functions. Subsequently, Dasgupta builds an algorithm for interactive learning using a generic supervised learning algorithm; feature adaptation using clustering; selective presentation to the human user and recording feedback on cluster features and cluster inter-relationships; and a cost function or loss function that could be optimized using suggested data structures. He illustrates his approach through elaborations on hierarchical clustering and an example from an animal classification tree.

It is an interesting talk that introduces the topic well and presents the case for the need to develop an interactive learning protocol very well. However, I could not be convinced about the framework and methodology suggested and feel further “learning” is required. The works of many fellow researchers are acknowledged in the course of the talk. However, no references are given.

This talk could be used to introduce a graduate student to the field of interactive machine learning.