An FPGA distributed implementation model for embedded SOM with on-line learning

Neuroengineering has contributed to the increasing capability of embedded hardware to efficiently execute neural computation models. These hardware structures, suitable for implementing only the recall phase or both phases - learning and recall - of artificial neural networks (ANNs) are generally identified as neuromorphic systems. Some features of the FPGA technology resemble characteristics usually associated to ANNs, such as the parallel processing and the configurability of digital circuits. Embedded neural systems elaborated to execute Self-Organizing Maps (SOMs) can take advantage of the circuit parallelism to speed-up training process and also of the design flexibility by the hardware reconfiguration enabled by FPGA chips. Hardware-based SOMs capable of executing the training process on chip are relevant because they can be employed in a wide variety of stand-alone applications. loT and mobile robotics are examples of some challenging research fields that use unsupervised learning to extract information from the surrounding environment and require the portability of embedded systems. This work presents a distributed architecture to implement SOM learning and recall algorithms into FPGAs. The distributed model does not employ any type of central control unit hence allowing a simple hardware redesign process during the configuration of the neural system. The set of experiments used for verifying the proposed architecture is specified in a data stream format. The aim of choosing streaming condition is to simulate applications in which the FPGA-based SOM must process a continuous flow of information autonomously. All the steps of the SOM learning algorithm to process an input sample are performed in a single clock cycle and implementation results show a maximum clock speed of 2.38 MHz. Synthesis reports on FPGA implementation of the distributed architecture for the experiments carried out herein also indicate that the factor of increase in consumption of chip resources is lower than the increment in the total number of SOM connections.