Discrete Intelligible Recognition Network

We present a new approach to recognize object and we test it in MNIST data set. The main purpose of this method is to solve some problems encountered by most current artificial intelligence. Firstly, most current artificial intelligence is incomprehensible. Second, most AI algorithm use a large number of additions and multiplications, but in our brain, it will be difficult to implement or learn multipliers automatically, and gradient descent is also difficult to achieve. In our method, we hope that we can implement a kind of algorithm that does not use gradient descent and may not use adders and multipliers and the algorithm is comprehensible. We propose a new method. The method does not use multipliers and adders as much as possible. For intelligibility, we discretize the parameters of the entire algorithm. The parameters can be regarded as an encoding. Algorithm recognizes objects by means of comparators. It first needs to remember the identified object and decompose this object into smaller objects through continuous learning, and grasp more changes of objects by creating more branches. Because the algorithm is identified by comparing similarities, we can learn how the object is identified by the path of the algorithm's execution, and thus understand how the algorithm makes decisions. You can get the code from: https://github.com/msq17/Discrete-Intelligible-Recognition-Network