Karim Hocine

Karim Hocine

Auto-generated profile, based on publicly available information

Karim Hocine

Sorbonne Université, CNRS, LIP6, Paris, France

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
26 September 2024 in ACM Transactions on Embedded Computing Systems
Verified
Journal Article
Open access
TGThomas Garbay
Thomas Garbay
KHKhalil Hachicha
Khalil Hachicha
PDPetr Dobiáš
Petr Dobiáš
APAndrea Pinna
Andrea Pinna
KHKarim Hocine
Karim Hocine
BGBertrand Granado
Bertrand Granado
Embedded systems based on Microcontroller Units (MCUs) often gather significant quantities of data and solve various issues. Convolutional Neural Networks (CNNs) have proven their effectiveness in solving computer vision and natural language processing tasks. However, implementing CNNs within MCUs is challenging due to their high inference costs, which varies widely depending on hardware targets and CNN topologies. Despite state-of-the-art advancements, no efficient design space exploration solutions handle the wide variety of implementation solutions. In this article, we introduce the ZIP-CNN design space exploration methodology, which facilitates CNN implementation within MCUs. We developed a model that quantitatively estimates the latency, energy consumption, and memory space required to run a CNN within an MCU. This model accounts for algorithmic reductions such as knowledge distillation, pruning, or quantization and applies to any CNN topology. To demonstrate the efficiency of our methodology, we investigated LeNet5, ResNet8, and ResNet26 within three different MCUs. We made materials and supplementary results available in a GitHub repository: https://github.com/ThGbay/ZIP-CNN . The proposed method was empirically verified on three hardware targets running at 14 different operating frequencies. The three CNN topologies investigated were implemented in their default configuration in FP32, and also reduced with INT8 quantization, pruning at five different rates and with knowledge distillation. The estimates of our model are very reliable with an error of 3.29% to 15.23% for latency, 3.12% to 10.34% for energy consumption, and 1.95% to 6.31% for memory space. These results are based on on-device measurements.