Hardware Accelerators for Evolving Building Block Modules for Artificial Brains

Abstract

Summary form only given. This paper argues that it is technologically possible to build artificial brains at relatively low cost. The proposed approach to doing this is to evolve large numbers (tens of thousands) of neural network modules, each with its own simple function, and then interconnect them inside a computer that would execute the neural signaling of the whole brain in real time, performing functions such as controlling the behaviors of a robot. The modules could be configured automatically using evolutionary algorithms, by a successive reconfiguration on field programmable gate arrays (FPGA), placed on commercially available boards such as those offered by Celoxica. These chips could be programmed using high level languages, such as "Handel-C", whose statements are "hardware compiled" into the chip configuring instructions to wire up the chip, speeding-up the execution of instructions. The major challenge of this approach is architecting the artificial brain - how to put 10,000s of evolved neural net modules together to perform a library of controllable behaviors. One potential concern of this approach relates to the anticipated unwanted synergy of inter module neural signaling. While most current artificial brain projects use supercomputers or PC clusters with 1000s of nodes, Moore's law facilitates increasingly larger computational power at low costs, making brain building technically and economically possible. Examples from our efforts in evolving neural modules are presented, along with a critical analysis of the state of the art and realistic assessment of the challenges ahead