A technique in machine learning and artificial intelligence that involves training models using gradient-based optimization. Differentiable programs are programs that rewrite themselves at least one component by optimizing along a gradient like neural networks do use optimization algorithms such as gradient descent.