FYI, if you analytically determine the recurrent connections, you get way better results than setting the connections randomly. http://compneuro.uwaterloo.ca/publications/voelker2018.html