System Identification with General Dynamic Neural Networks and Network Pruning

This paper presents an exact pruning algorithm with adaptive pruning interval for general dynamic neural networks (GDNN). GDNNs are artificial neural networks with internal dynamics. All layers have feedback connections with time delays to the same and to all other layers. The structure of the plant is unknown, so the identification process is started with a larger network architecture than necessary. During parameter optimization with the Levenberg- Marquardt (LM) algorithm irrelevant weights of the dynamic neural network are deleted in order to find a model for the plant as simple as possible. The weights to be pruned are found by direct evaluation of the training data within a sliding time window. The influence of pruning on the identification system depends on the network architecture at pruning time and the selected weight to be deleted. As the architecture of the model is changed drastically during the identification and pruning process, it is suggested to adapt the pruning interval online. Two system identification examples show the architecture selection ability of the proposed pruning approach.

Performance Evaluation of Neural Network Prediction for Data Prefetching in Embedded Applications

Embedded systems need to respect stringent real time constraints. Various hardware components included in such systems such as cache memories exhibit variability and therefore affect execution time. Indeed, a cache memory access from an embedded microprocessor might result in a cache hit where the data is available or a cache miss and the data need to be fetched with an additional delay from an external memory. It is therefore highly desirable to predict future memory accesses during execution in order to appropriately prefetch data without incurring delays. In this paper, we evaluate the potential of several artificial neural networks for the prediction of instruction memory addresses. Neural network have the potential to tackle the nonlinear behavior observed in memory accesses during program execution and their demonstrated numerous hardware implementation emphasize this choice over traditional forecasting techniques for their inclusion in embedded systems. However, embedded applications execute millions of instructions and therefore millions of addresses to be predicted. This very challenging problem of neural network based prediction of large time series is approached in this paper by evaluating various neural network architectures based on the recurrent neural network paradigm with pre-processing based on the Self Organizing Map (SOM) classification technique.