While Convolutional Neural Networks (CNNs) are going mainstream and deployed in widespread domains including safety-critical systems, the impact of reliability issues in CNN-hosting systems are still under-explored. In this paper, we experimentally evaluate the inherent fault tolerance of CNNs through fault injection experiments. Our experiments demonstrate that quantization increases reliability. We also show the most vulnerable bits in the IEEE 754 format representation. At the layer-level, a study on the network architecture was performed to identify the most vulnerable layers. This exploratory study is useful for comprehensive critical system reliability enhancement since only the vulnerable parts will be protected.

Driver behaviors and decisions are crucial factors for on-road driving safety. With a precise driver behavior monitoring system, traffic accidents and injuries can be significantly reduced. However, understanding human behaviors in real-world driving settings is a challenging task because of the uncontrolled conditions including illumination variation, occlusion, and dynamic and cluttered background. In this paper, a Kinect sensor, which provides multimodal signals, is adopted as a driver monitoring sensor to recognize safe driving and common secondary most distracting in-vehicle actions. We propose a novel soft spatial attention-based network named the Depth-based Spatial Attention network (DSA), which adds a cognitive process to deep network by selectively focusing on the driver’s silhouette and motion in the cluttered driving scene. In fact, at each time t, we introduce a new weighted RGB frame based on an attention model designed using a depth frame. The final classification accuracy is substantially enhanced compared to the state-of-the-art results with an achieved improvement of up to 27%.

Approximate computing is a design paradigm considered for a range of applications that can tolerate some loss of accuracy. In fact, the bottleneck in conventional digital design techniques can be eliminated to achieve higher performance and energy efficiency by compromising accuracy. In this paper, a new architecture that engages accuracy as a design parameter is presented where an approximate parallel multiplier using heterogeneous blocks is implemented. Based on design space exploration, we demonstrate that introducing diverse building blocks to implement the multiplier rather than cloning one building block achieves higher precision results. We show experimental results in terms of precision, delay and power dissipation as metrics and compare with 3 previous approximate designs. Our results show that the proposed heterogeneous multiplier achieves more precise outputs than the tested circuits while improving performance and power tradeoffs.