Mohsen Raji

Mohammadamin Baghbanbashi, Mohsen Raji, Behnam Ghavami

YOLO is a deep neural network (DNN) model presented for robust real-time object detection following the one-stage inference approach. It outperforms other real-time object detectors in terms of speed and accuracy by a wide margin. Nevertheless, since YOLO is developed upon a DNN backbone with numerous parameters, it will cause excessive memory load, thereby deploying it on memory-constrained devices is a severe challenge in practice. To overcome this limitation, model compression techniques, such as quantizing parameters to lower-precision values, can be adopted. As the most recent version of YOLO, YOLOv7 achieves such state-of-the-art performance in speed and accuracy in the range of 5 FPS to 160 FPS that it surpasses all former versions of YOLO and other existing models in this regard. So far, the robustness of several quantization schemes has been evaluated on older versions of YOLO. These methods may not necessarily yield similar results for YOLOv7 as it utilizes a different architecture. In this paper, we conduct in-depth research on the effectiveness of a variety of quantization schemes on the pre-trained weights of the state-of-the-art YOLOv7 model. Experimental results demonstrate that using 4-bit quantization coupled with the combination of different granularities results in ~3.92x and ~3.86x memory-saving for uniform and non-uniform quantization, respectively, with only 2.5% and 1% accuracy loss compared to the full-precision baseline model.

Kimia Soroush, Mohsen Raji, Behnam Ghavami

Deep neural networks (DNNs) have demonstrated remarkable performance in many tasks but it often comes at a high computational cost and memory usage. Compression techniques, such as pruning and quantization, are applied to reduce the memory footprint of DNNs and make it possible to accommodate them on resource-constrained edge devices. Recently, explainable artificial intelligence (XAI) methods have been introduced with the purpose of understanding and explaining AI methods. XAI can be utilized to get to know the inner functioning of DNNs, such as the importance of different neurons and features in the overall performance of DNNs. In this paper, a novel DNN compression approach using XAI is proposed to efficiently reduce the DNN model size with negligible accuracy loss. In the proposed approach, the importance score of DNN parameters (i.e. weights) are computed using a gradient-based XAI technique called Layer-wise Relevance Propagation (LRP). Then, the scores are used to compress the DNN as follows: 1) the parameters with the negative or zero importance scores are pruned and removed from the model, 2) mixed-precision quantization is applied to quantize the weights with higher/lower score with higher/lower number of bits. The experimental results show that, the proposed compression approach reduces the model size by 64% while the accuracy is improved by 42% compared to the state-of-the-art XAI-based compression method.

Fereshteh Baradaran, Mohsen Raji, Azadeh Baradaran et al.

Vision Transformers (ViTs) have demonstrated superior performance over Convolutional Neural Networks (CNNs) in various vision-related tasks such as classification, object detection, and segmentation due to their use of self-attention mechanisms. As ViTs become more popular in safety-critical applications like autonomous driving, ensuring their correct functionality becomes essential, especially in the presence of bit-flip faults in their parameters stored in memory. In this paper, a fault tolerance technique is introduced to protect ViT parameters against bit-flip faults with zero memory overhead. Since the least significant bits of parameters are not critical for model accuracy, replacing the LSB with a parity bit provides an error detection mechanism without imposing any overhead on the model. When faults are detected, affected parameters are masked by zeroing out, as most parameters in ViT models are near zero, effectively preventing accuracy degradation. This approach enhances reliability across ViT models, improving the robustness of parameters to bit-flips by up to three orders of magnitude, making it an effective zero-overhead solution for fault tolerance in critical applications.

Kimia Soroush, Nastaran Shirazi, Mohsen Raji

Deep Neural Networks (DNNs) are widely employed in safety-critical domains, where ensuring their reliability is essential. Triple Modular Redundancy (TMR) is an effective technique to enhance the reliability of DNNs in the presence of bit-flip faults. In order to handle the significant overhead of TMR, it is applied selectively on the parameters and components with the highest contribution at the model output. Hence, the accuracy of the selection criterion plays the key role on the efficiency of TMR. This paper presents an efficient TMR approach to enhance the reliability of DNNs against bit-flip faults using an Explainable Artificial Intelligence (XAI) method. Since XAI can provide valuable insights about the importance of individual neurons and weights in the performance of the network, they can be applied as the selection metric in TMR techniques. The proposed method utilizes a low-cost, gradient-based XAI technique known as Layer-wise Relevance Propagation (LRP) to calculate importance scores for DNN parameters. These scores are then used to enhance the reliability of the model, with the most critical weights being protected by TMR. The proposed approach is evaluated on two DNN models, VGG16 and AlexNet, using datasets such as MNIST and CIFAR-10. The results demonstrate that the method can protect the AlexNet model at a bit error rate of 10-4, achieving over 60% reliability improvement while maintaining the same overhead as state-of-the-art methods.