Information theoretic learning for pixel-based visual agents