PERTINENCE: Input-based Opportunistic Neural Network Dynamic Execution, AxMoE: Characterizing the Impact of Approximate Multipliers on Mixture-of-Experts DNN Architectures, Flip-MoE: Characterizing Fault Resilience of Mixture-of-Experts Architectures under Weight-Level Bit Flips, 06\/06\/2025<\/strong> – Scientific seminar<\/strong> at Inria Center at Rennes University by Gayathri Ananthanarayanan<\/strong>, ass. professor, Department of Computer Science and Engineering of the Indian Institute of Technology (IIT) Dharwad.
Omkar Shende, Gayathri Ananthanarayanan, and Marcello Traiola
IEEE Access, June, 2026.
DOI: 10.1109\/ACCESS.2026.3707342
URL: https:\/\/arxiv.org\/abs\/2507.01695v2<\/a><\/p>\n\n\n\n
Omkar B Shende, Marcello Traiola, and Gayathri Ananthanarayanan
In IEEE Computer Society Annual Symposium on VLSI (ISVLSI) 2026, July, 2026.
URL: https:\/\/arxiv.org\/abs\/2605.04754<\/a><\/p>\n\n\n\n
Omkar Shende, Marcello Traiola, and Gayathri Ananthanarayanan
In 2026 IEEE International Symposium on Defect and Fault Tolerance in VLSI and Nanotechnology Systems (DFT), October, 2026<\/a><\/p>\n\n\n\n
\n\n\n\nEvents and scientific dissemination<\/h2>\n\n\n\n
“Towards Efficient, Adaptive, and Dynamic Neural Inference on Heterogeneous Platforms”<\/strong>
Abstract: <\/strong>Deploying deep neural networks (DNNs) on edge devices presents unique challenges due to limited computational resources, diverse hardware architectures, and dynamic runtime conditions. This talk presents a comprehensive overview of our research efforts aimed at making DNN inference on heterogeneous edge platforms more efficient, adaptive, and responsive to real-world constraints.
The talk will begin with the introduction of a throughput-oriented scheduling framework tailored for ARM big.LITTLE processors, which exploits pipeline parallelism to maximize inference performance. Next, I will discuss a co-execution-aware framework that intelligently maps multiple DNNs to various heterogeneous accelerators based on user-defined trade-offs between power, throughput, and accuracy. The talk will also introduce PERTINENCE, an input-aware inference strategy that dynamically selects the most suitable model from a pool of pre-trained networks based on input complexity, achieving significant computational savings without compromising accuracy.<\/p>\n\n\n\n