In this paper, the authors report a way to use concepts from statistical learning to gain an advantage in terms of error exponents while communicating over a discrete memoryless channel. The study utilizes the simulation capability of the scientific computing package MATLAB to show that the proposed decoding method performs better than the traditional method of joint typicality decoding. The advantage is secured by modifying the traditional specification of what constitutes a decoding error. This is justified by the paradigm, also used in the program of `utilizing' noisy feedback, that one ought not to declare a condition as an error if some further processing can extract useful information from it.