Abstract:
Differentiable programming is able to combine different functions or modules in a data processing pipeline with the goal of applying gradient descent-based end-to-end learning or optimization. A significant impediment to differentiable programming is the non-differentiable nature of some functions. We propose to overcome this difficulty by using neural networks to approximate such modules. An approximating neural network provides synthetic gradients (SG) for backpropagation across a non-differentiable module. Our design is grounded on a well-known theory that gradient of an approximating neural network can approximate a sub-gradient of a weakly differentiable function. We apply SG to combine convolutional neural network (CNN) with dynamic programming (DP) in end-to-end learning for segmenting left ventricle from short axis view of heart MRI. Our experiments show that end-to-end combination of CNN and DP requires fewer labeled images to achieve a significantly better segmentation accuracy than using only CNN.
