Hardware simulator for a deep belief netrowk system

Deep belief netrowk hardware simulated on MATLAB

This is a register-transfer-level (RTL) hardware simulator implemented in MATLAB simulating a special-purpose hardware system for deep belief networks (DBNs). This is a part of supplementary information for a paper titled "Computation of deep belief networks using special-purpose hardware architecture." published at IEEE IJCNN 2014 conference. http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=6889903
In machine learning, DBN is a generative graphical model, or alternatively a type of deep neural network, composed of multiple layers of restricted Boltzmann machines.
This hardware system is based on what we call Neuron Machine (NM) hardware architecture that can be used specifically for neural network systems. Through this simulation, along with other matrials, we are trying to explain how our system can achieve a performance-resource ratio several orders of magnitude better than CPUs and GPUs, impling smaller chip area and power consumption, and better computational speed using the same resource. Even though the source code is coded in a sequential language, it is cycle-accurate and simulates various RTL components such as registers, arithmetic operators and memories, in the same way as the hardware. Readers can execute the simulator on the fly and monitor how it is working.
For more on this hardware simulator, refer a web site explaining it. http://neurocomputings.com/papers/dbn
Each file runs stand alone. By default, it trains a DBN with a set of characters and shows mean-square error, which is a distance between the training data and what the system learned, is decreasing. The file nm_dbn_oputil.m and nm_dbn_mem.m executes the same simulation showing the utilization of arithmetic operators and memories, respectively.