Guest_Jim_* Posted April 16, 2015 Posted April 16, 2015 One of the many uses of machine learning is computer vision, whereby a computer analyzes a scene and identifies the objects in it, without them exactly matching known models. Building these algorithms can requires thousands of lines of code, but some are turning to probabilistic programming languages to simplify the work. In the case of the Picture language MIT researchers developed, those thousands of lines can be simplified to less than 50. Probabilistic programming differs from deterministic programming by being based more on inference, which fits well with machine learning. Instead of requiring very specific descriptions, programmers can describe a vague model that the program runs through inference schemes to solve, using inverse-graphics reasoning. The MIT researchers have tested Picture by giving it the simple description of the human face as having two symmetrically placed eyes, with the nose and mouth positioned beneath them. Armed just with that knowledge and examples, the program was tasked with working through 2D images to construct 3D models, and was able to match the thousand-line programs, and in some cases surpass them. Technically more code than 50 lines is involved behind Picture, as it draws on multiple inference algorithms, but the model for the task itself is still much simpler. This is actually one of the purposes of probabilistic programming languages, where the language is generic and can be used with many inference algorithms, depending on the task. Source: MIT Share this post Link to post Share on other sites More sharing options...
Recommended Posts