Inverse Diffusion by Proximal Optimization with TensorFlow

Abstract: We present a method for implementing a large scale, proximal optimization algorithm in a machine learning framework to solve an inverse problem. The algorithm is based on a previously developed method for analyzing data of some imagebased immunoassays in the context of detecting diffused cells. By employing TensorFlow through it’s Python API, parallelized computations on graphical processing units, distributed processing, automatic gradient computation and computational efficiency are made available for implementation. Image processing methods are also utilized throughout the implementation, resulting in a F1-score of 0.91. The performance matches previous implementations of the algorithm presented, leading to the conclusion that the TensorFlow platform is well suited and provides advantages over traditional methods. We conclude that the method presented coupled with the resulting performance is a proof of concept, by providing an example of implementing convex optimization in a machine learning framework.