Information Theoretic Learning for Pixel-Based Visual Agents