Recurrent Neural Networks are ubiquitous and pervasive in many artificial intelligence applications such as speech recognition, predictive healthcare, creative art, and so on. Although they provide accurate superior solutions, they pose a massive challenge “training havoc.” Current expansion of IoT demands intelligent models to be deployed at the edge. This is precisely to handle increasing model sizes and complex network architectures. Design efforts to meet these for greater performance have had inverse effects on portability on edge devices with real-time constraints of memory, latency, and energy. This article provides a detailed insight into various compression techniques widely disseminated in the deep learning regime. They have become key in mapping powerful RNNs onto resource-constrained devices. While compression of RNNs is the main focus of the survey, it also highlights challenges encountered while training. The training procedure directly influences model performance and compression alongside. Recent advancements to overcome the training challenges with their strengths and drawbacks are discussed. In short, the survey covers the three-step process, namely, architecture selection, efficient training process, and suitable compression technique applicable to a resource-constrained environment. It is thus one of the comprehensive survey guides a developer can adapt for a time-series problem context and an RNN solution for the edge.
The Recurrent Neural Networks (RNNs) are an essential class of supervised learning algorithms. Complex tasks like speech recognition, machine translation, sentiment classification, weather prediction, etc., are now performed by well-trained RNNs. Local or cloud-based GPU machines are used to train them. However, inference is now shifting to miniature, mobile, IoT devices and even micro-controllers. Due to their colossal memory and computing requirements, mapping RNNs directly onto resource-constrained platforms is arcane and challenging. The efficacy of edge-intelligent RNNs (EI-RNNs) must satisfy both performance and memory-fitting requirements at the same time without compromising one for the other. This study’s aim was to provide an empirical evaluation and optimization of historic as well as recent RNN architectures for high-performance and low-memory footprint goals. We focused on Human Activity Recognition (HAR) tasks based on wearable sensor data for embedded healthcare applications. We evaluated and optimized six different recurrent units, namely Vanilla RNNs, Long Short-Term Memory (LSTM) units, Gated Recurrent Units (GRUs), Fast Gated Recurrent Neural Networks (FGRNNs), Fast Recurrent Neural Networks (FRNNs), and Unitary Gated Recurrent Neural Networks (UGRNNs) on eight publicly available time-series HAR datasets. We used the hold-out and cross-validation protocols for training the RNNs. We used low-rank parameterization, iterative hard thresholding, and spare retraining compression for RNNs. We found that efficient training (i.e., dataset handling and preprocessing procedures, hyperparameter tuning, and so on, and suitable compression methods (like low-rank parameterization and iterative pruning) are critical in optimizing RNNs for performance and memory efficiency. We implemented the inference of the optimized models on Raspberry Pi.