This paper introduces a novel approach to human-machine collaborative learning that allows for the chronically missing human learnability in the context of supervised machine learning. The basic tenet of this approach is the refinement of a human designed software model through the iterative learning loop. Each iteration of the loop consists of two phases: (i) automatic data-driven parameter adjustment, performed by means of stochastic greedy local search, and (ii) human-driven model adjustment based on insights gained in the previous phase. The proposed approach is demonstrated through a real-life study of automatic electricity meter reading in the presence of noise. Thus, a cognitively-inspired non-connectionist approach to digit detection and recognition is introduced, which is subject to refinement through the iterative process of human-machine cooperation. The prototype system is evaluated with respect to the recognition accuracy (with the highest digit recognition accuracy of 94%), and also discussed with respect to the storage requirements, generalizability, utilized contextual information, and efficiency.