The CHIR learning algorithm: new results and improvements

Learning by Choice of Internal Representations (CHIR) is a training method for feedforward networks of binary units. This paper gives a short description of the algorithm and presents some improvements relevant for practical application `real life´ problems. It demonstrates the importance of changing the internal representation during learning, and the insensitivity of the algorithm to the choice of parameters. Ability to train networks on analog input and output patterns is discussed