Deep Active Audio Feature Learning in Resource-Constrained Environments