@misc{TaheriPatneCherezovaetal., author = {Taheri, Mahdi and Patne, Parth and Cherezova, Natalia and Mahani, Ali and Herglotz, Christian and Jenihhin, Maksim}, title = {RL-agent-based early-exit DNN architecture search framework}, series = {IEEE 28th International Symposium on Design and Diagnostics of Electronic Circuits and Systems (DDECS) : proceedings}, journal = {IEEE 28th International Symposium on Design and Diagnostics of Electronic Circuits and Systems (DDECS) : proceedings}, publisher = {IEEE}, address = {Piscataway, NJ}, isbn = {979-8-3315-2801-0}, issn = {2334-3133}, doi = {10.1109/DDECS63720.2025.11006795}, pages = {145 -- 148}, abstract = {This paper introduces a Reinforcement Learning (RL)-based framework for optimizing early-exit configurations in Deep Neural Networks (DNNs). By integrating RL with BranchyNet-inspired architectures, the framework dynamically determines optimal early exit placements and confidence thresholds, balancing inference time, energy consumption, and accuracy. Key contributions include an early-exit DNN architecture search, an RL-driven threshold optimization process during training, and a design-space exploration open-source framework. Experiments on models such as ResNet-18, VGG-16, and AlexNet, using benchmarks like CIFAR-10 and MNIST, reveal significant reductions in inference time (up to 69.7x) and power consumption while keeping accuracy drop within 1-2\%. This work demonstrates that dynamic early-exit strategies can enhance DNN efficiency while maintaining performance, paving the way for resource-constrained applications.}, language = {en} } @misc{NazariTaheriAzarpeyvandetal., author = {Nazari, Samira and Taheri, Mahdi and Azarpeyvand, Ali and Afsharchi, Mohsen and Herglotz, Christian and Jenihhin, Maksim}, title = {GENIE : GENetIc algorithm-based REliability assessment methodology for deep neural networks}, series = {2025 11th International Conference on Computing and Artificial Intelligence (ICCAI)}, journal = {2025 11th International Conference on Computing and Artificial Intelligence (ICCAI)}, publisher = {IEEE}, address = {Piscataway, NJ}, isbn = {979-8-3315-2492-0}, doi = {10.1109/ICCAI66501.2025.00049}, pages = {1 -- 8}, abstract = {As deep neural networks (DNNs) are becoming vital to numerous safety-critical applications, ensuring their fault reliability is crucial. This paper presents a hybrid framework that combines genetic algorithms with fault injection and analytical methods to identify vulnerable neurons and layers in DNNs. Validated on LeNet-5, AlexNet, and VGG-11, our approach significantly enhances model accuracy in harsh environments by protecting critical components, achieving reliability improvements (accuracy drop of the model compared to the golden network after fault injection) of 69.86\% for LeNet, and 99.65\% for AlexNet at BER=1E-4, and 91.37\% improvement for VGG at BER=1E-6. Furthermore, our framework reduces computational costs by requiring fewer inferences than traditional analytical methods, highlighting its potential to improve DNN accelerators' reliability and contribute to their safety.}, language = {en} } @misc{SharifianTaheriRashtchietal., author = {Sharifian, Saeed and Taheri, Mahdi and Rashtchi, Vahid and Azarpeyvand, Ali and Herglotz, Christian and Jenihhin, Maksim}, title = {Reliability-aware hyperparameter optimization for ANN-to-SNN conversion}, series = {WiPiEC journal : works in progress in embedded computing journal : special issue}, volume = {11}, journal = {WiPiEC journal : works in progress in embedded computing journal : special issue}, number = {1}, publisher = {WiPiEC Journal}, address = {Podgorica, Montenegro}, issn = {2980-7298}, doi = {https://doi.org/10.64552/wipiec.v11i1.85}, pages = {1 -- 7}, abstract = {Spiking Neural Networks (SNNs) have emerged as an energy-efficient alternative to Artificial Neural Networks (ANNs), particularly for edge-computing and safety-critical applications. Unlike conventional ANNs, SNNs leverage sparse event-driven processing to reduce energy consumption while significantly maintaining high computational efficiency. This paper presents a framework designed to optimize the conversion of ANNs into equivalent SNNs while balancing accuracy, reliability, and energy efficiency. The proposed framework systematically explores SNN hyperparameters to identify configurations that achieve superior performance compared to their ANN counterparts. Experimental evaluations on MNIST and Fashion-MNIST datasets with different network topologies demonstrate that the optimized SNNs achieve comparable accuracy while offering in some cases 27.81× and 15.17× lower energy consumption and 1.92× and 1.84× less accuracy drop in the presence of faults, respectively, over the ANN baseline. The results highlight the applicability of SNNs in reliability-critical power-constrained environments.}, language = {en} } @misc{PappalardoBellarminoDeveautouretal., author = {Pappalardo, Salvatore and Bellarmino, Nicolo' and Deveautour, Bastien and Bosio, Alberto and Taheri, Mahdi and Daneshtalab, Masoud and Raik, Jaan and Jenihhin, Maksim}, title = {SAFFIRA : a framework for assessing the reliability of systolic-array DNN accelerators}, series = {Journal of circuits, systems, and computers : JCSC}, volume = {34}, journal = {Journal of circuits, systems, and computers : JCSC}, number = {18}, publisher = {World Scientific}, address = {Singapore}, issn = {1793-6454}, doi = {https://doi.org/10.1142/S0218126625430017}, abstract = {Systolic array has emerged as a prominent architecture for Deep Neural Network (DNN) hardware accelerators, providing high-throughput and low-latency performance essential for deploying DNNs across diverse applications. However, when used in safety-critical applications, reliability assessment is mandatory to guarantee the correct behavior of DNN accelerators. While fault injection stands out as a well-established practical and robust method for reliability assessment, it is still a very time-consuming process. This paper addresses the time efficiency issue by introducing a novel hierarchical software-based hardware-aware fault injection strategy tailored for systolic array-based DNN accelerators. A Uniform Recurrent Equations system is used for software modeling of the systolic-array core of the DNN accelerators. The approach demonstrates a reduction of the fault injection time up to 3× compared to the state-of-the-art hybrid (software/hardware) hardware-aware fault injection frameworks and more than 2000× compared to RT-level fault injection frameworks without compromising the accuracy from the application level. Additionally, we introduce novel reliability metrics to better evaluate the robustness of a deep neural network system. The performance of the framework is studied on state-of-the-art DNN benchmarks.}, language = {en} } @misc{ZhangVassiliadisDorneanuetal., author = {Zhang, Sushen and Vassiliadis, Vassilios S. and Dorneanu, Bogdan and Arellano-Garc{\´i}a, Harvey}, title = {Hierarchical multi-scale parametric optimization of deep neural networks}, series = {Applied Intelligence}, volume = {53}, journal = {Applied Intelligence}, number = {21}, issn = {1573-7497}, doi = {10.1007/s10489-023-04745-8}, pages = {24963 -- 24990}, abstract = {Traditionally, sensitivity analysis has been utilized to determine the importance of input variables to a deep neural network (DNN). However, the quantification of sensitivity for each neuron in a network presents a significant challenge. In this article, a selective method for calculating neuron sensitivity in layers of neurons concerning network output is proposed. This approach incorporates scaling factors that facilitate the evaluation and comparison of neuron importance. Additionally, a hierarchical multi-scale optimization framework is proposed, where layers with high-importance neurons are selectively optimized. Unlike the traditional backpropagation method that optimizes the whole network at once, this alternative approach focuses on optimizing the more important layers. This paper provides fundamental theoretical analysis and motivating case study results for the proposed neural network treatment. The framework is shown to be effective in network optimization when applied to simulated and UCI Machine Learning Repository datasets. This alternative training generates local minima close to or even better than those obtained with the backpropagation method, utilizing the same starting points for comparative purposes within a multi-start optimization procedure. Moreover, the proposed approach is observed to be more efficient for large-scale DNNs. These results validate the proposed algorithmic framework as a rigorous and robust new optimization methodology for training (fitting) neural networks to input/output data series of any given system.}, language = {en} }