Making most of quite little: Improving AI training for edge