Kaoutar El Maghraoui
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Kaoutar El Maghraoui has a total of 74 co-authors. Here are the top co-authors by number of shared publications:
- Hadjer Benmeziane
IBM Research Europe, Rüschlikon, Switzerland
5 shared publications - Liu Liu
Rensselaer Polytechnic Institute
3 shared publications - Smail Niar
Université Polytechnique des Hauts-de-France, Valenciennces, France
3 shared publications Abu Sebastian
IBM Research Europe, Rüschlikon, Switzerland
2 shared publicationsHadjer Benmeziane
IBM Research Europe
2 shared publications- Hamza Ouarnoughi
Higher National School of Computer Science -ESI ex INI Université Polytechnique Hauts-de-France LAMIH CNRS Smail Niar Université Polytechnique Hauts-de-France LAMIH/CNRS INSA Hauts-de-France
2 shared publications
Latest Publications
- 29 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal on Selected Areas in Information Theory
- Vol. 7, 279-289
On the Convergence Theory of Pipeline Gradient-Based Analog In-Memory Training
Aiming to accelerate the training of large deep neural networks (DNN) in an energy-efficient way, analog inmemory computing (AIMC) emerges as a solution with immense potential. AIMC accelerator keeps model weights in memory without moving them from memory to processors during training, reducing overhead dramatically. Despite its efficiency, scaling up AIMC systems presents significant challenges. Since weight copying is expensive and inaccurate, data parallelism is less efficient on AIMC accelerators. It necessitates the exploration of pipeline parallelism, particularly asynchronous pipeline parallelism, which utilizes all available accelerators during the training process. This paper examines the convergence theory of stochastic gradient descent on AIMC hardware with an asynchronous pipeline (Analog-SGD-AP). Although there is empirical exploration of AIMC accelerators, the theoretical understanding of how analog hardware imperfections in weight updates affect the training of multi-layer DNN models remains underexplored. Furthermore, the asynchronous pipeline parallelism results in stale weights issues, which render the update signals no longer valid gradients. To close the gap, this paper investigates the convergence properties of Analog-SGD-AP on multi-layer DNN training. We show that the Analog-SGD-AP converges with iteration complexity O (ε–2+ε–1) despite the aforementioned issues, which matches the complexities of digital SGD and Analog SGD with synchronous pipeline, except the non-dominant term O(ε–1). It implies that AIMC training benefits from asynchronous pipelining almost for free compared with the synchronous pipeline by overlapping computation.
- 29 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal on Selected Areas in Information Theory
- Vol. 7, 279-289
On the Convergence Theory of Pipeline Gradient-Based Analog In-Memory Training
Aiming to accelerate the training of large deep neural networks (DNN) in an energy-efficient way, analog inmemory computing (AIMC) emerges as a solution with immense potential. AIMC accelerator keeps model weights in memory without moving them from memory to processors during training, reducing overhead dramatically. Despite its efficiency, scaling up AIMC systems presents significant challenges. Since weight copying is expensive and inaccurate, data parallelism is less efficient on AIMC accelerators. It necessitates the exploration of pipeline parallelism, particularly asynchronous pipeline parallelism, which utilizes all available accelerators during the training process. This paper examines the convergence theory of stochastic gradient descent on AIMC hardware with an asynchronous pipeline (Analog-SGD-AP). Although there is empirical exploration of AIMC accelerators, the theoretical understanding of how analog hardware imperfections in weight updates affect the training of multi-layer DNN models remains underexplored. Furthermore, the asynchronous pipeline parallelism results in stale weights issues, which render the update signals no longer valid gradients. To close the gap, this paper investigates the convergence properties of Analog-SGD-AP on multi-layer DNN training. We show that the Analog-SGD-AP converges with iteration complexity O (ε–2+ε–1) despite the aforementioned issues, which matches the complexities of digital SGD and Analog SGD with synchronous pipeline, except the non-dominant term O(ε–1). It implies that AIMC training benefits from asynchronous pipelining almost for free compared with the synchronous pipeline by overlapping computation.
- Open Access
- 22 March 2026
- Published by Association for Computing Machinery (ACM)
- p. 1863-1879
STARC: Selective Token Access with Remapping and Clustering for Efficient LLM Decoding on PIM Systems
- 22 March 2026
- Published by Association for Computing Machinery (ACM)
- p. 1863-1879
Open AccessSTARC: Selective Token Access with Remapping and Clustering for Efficient LLM Decoding on PIM Systems
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Automation Science and Engineering
- Vol. 23, 18659-18676
SparseST: Exploiting Data Sparsity in Spatiotemporal Modeling and Prediction
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Automation Science and Engineering
- Vol. 23, 18659-18676
SparseST: Exploiting Data Sparsity in Spatiotemporal Modeling and Prediction
- 17 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Computer Architecture Letters
- Vol. 24 (2) , 337-340
Accelerating LLM Inference via Dynamic KV Cache Placement in Heterogeneous Memory System
Large Language Model (LLM) inference is increasingly constrained by memory bandwidth, with frequent access to the key-value (KV) cache dominating data movement. While attention sparsity reduces some memory traffic, the relevance of past tokens varies over time, requiring the full KV cache to remain accessible and sustaining pressure on both bandwidth and capacity. With advances in interconnects such as NVLink and LPDDR5X, modern AI hardware now integrates high-bandwidth memory (HBM) with high-speed off-package DRAM, making heterogeneous memory systems a practical solution. This work investigates dynamic KV cache placement across such systems to maximize aggregated bandwidth utilization under capacity constraints. Rather than proposing a specific scheduling policy, we formulate the placement problem mathematically and derive a theoretical upper bound, revealing substantial headroom for runtime optimization. To our knowledge, this is the first formal treatment of dynamic KV cache scheduling in heterogeneous memory systems for LLM inference.
- 17 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Computer Architecture Letters
- Vol. 24 (2) , 337-340
Accelerating LLM Inference via Dynamic KV Cache Placement in Heterogeneous Memory System
Large Language Model (LLM) inference is increasingly constrained by memory bandwidth, with frequent access to the key-value (KV) cache dominating data movement. While attention sparsity reduces some memory traffic, the relevance of past tokens varies over time, requiring the full KV cache to remain accessible and sustaining pressure on both bandwidth and capacity. With advances in interconnects such as NVLink and LPDDR5X, modern AI hardware now integrates high-bandwidth memory (HBM) with high-speed off-package DRAM, making heterogeneous memory systems a practical solution. This work investigates dynamic KV cache placement across such systems to maximize aggregated bandwidth utilization under capacity constraints. Rather than proposing a specific scheduling policy, we formulate the placement problem mathematically and derive a theoretical upper bound, revealing substantial headroom for runtime optimization. To our knowledge, this is the first formal treatment of dynamic KV cache scheduling in heterogeneous memory systems for LLM inference.
- Open Access
- 1 August 2024
- Published by International Joint Conferences on Artificial Intelligence
- Vol. 9, 7894-7903
Endowing Interpretability for Neural Cognitive Diagnosis by Efficient Kolmogorov-Arnold Networks
Cognitive diagnosis is crucial for intelligent education because of its ability to reveal students' proficiency in knowledge concepts. Although neural network-based neural cognitive diagnosis models (CDMs) have exhibited significantly better performance than traditional models, neural cognitive diagnosis is criticized for the poor model interpretability due to the multi-layer perceptron(MLP) employed, even with the monotonicity assumption. Therefore, this paper proposes to empower the interpretability of neural cognitive diagnosis models through efficient Kolmogorov-Arnold networks (KANs), named KAN2CD, where KANs are used to enhance interpretability in two manners. Specifically, in the first manner, KANs are directly used to replace the used MLPs in existing neural CDMs; while in the second manner, the student embedding, exercise embedding, and concept embedding are directly processed by several KANs, and then their outputs are further combined and learned in a unified KAN to get final predictions. Besides, the implementation of original KANs is modified without affecting the interpretability to overcome the problem of training KANs slowly. Extensive experiments show KAN2CD outperforms traditional CDMs and slightly surpasses existing neural CDMs, and its learned structures ensure interpretability on par with traditional CDMs and better than neural CDMs. The datasets, associated code, and more experimental results are available at https://github.com/null233QAQ/KAN2CD.
- 1 August 2024
- Published by International Joint Conferences on Artificial Intelligence
- Vol. 9, 7894-7903
Open AccessEndowing Interpretability for Neural Cognitive Diagnosis by Efficient Kolmogorov-Arnold Networks
Cognitive diagnosis is crucial for intelligent education because of its ability to reveal students' proficiency in knowledge concepts. Although neural network-based neural cognitive diagnosis models (CDMs) have exhibited significantly better performance than traditional models, neural cognitive diagnosis is criticized for the poor model interpretability due to the multi-layer perceptron(MLP) employed, even with the monotonicity assumption. Therefore, this paper proposes to empower the interpretability of neural cognitive diagnosis models through efficient Kolmogorov-Arnold networks (KANs), named KAN2CD, where KANs are used to enhance interpretability in two manners. Specifically, in the first manner, KANs are directly used to replace the used MLPs in existing neural CDMs; while in the second manner, the student embedding, exercise embedding, and concept embedding are directly processed by several KANs, and then their outputs are further combined and learned in a unified KAN to get final predictions. Besides, the implementation of original KANs is modified without affecting the interpretability to overcome the problem of training KANs slowly. Extensive experiments show KAN2CD outperforms traditional CDMs and slightly surpasses existing neural CDMs, and its learned structures ensure interpretability on par with traditional CDMs and better than neural CDMs. The datasets, associated code, and more experimental results are available at https://github.com/null233QAQ/KAN2CD.
- 20 May 2024
- Published by Springer Nature in Nature Reviews Electrical Engineering
- Vol. 1 (6) , 374-390
Neural architecture search for in-memory computing-based deep learning accelerators
- 20 May 2024
- Published by Springer Nature in Nature Reviews Electrical Engineering
- Vol. 1 (6) , 374-390
Neural architecture search for in-memory computing-based deep learning accelerators
- Open Access
Analog In-Memory Computing with Uncertainty Quantification for Efficient Edge-based Medical Imaging Segmentation
This work investigates the role of the emerging Analog In-memory computing (AIMC) paradigm in enabling Medical AI analysis and improving the certainty of these models at the edge. It contrasts AIMC's efficiency with traditional digital computing's limitations in power, speed, and scalability. Our comprehensive evaluation focuses on brain tumor analysis, spleen segmentation, and nuclei detection. The study highlights the superior robustness of isotropic architectures, which exhibit a minimal accuracy drop (0.04) in analog-aware training, compared to significant drops (up to 0.15) in pyramidal structures. Additionally, the paper emphasizes IMC's effective data pipelining, reducing latency and increasing throughput as well as the exploitation of inherent noise within AIMC, strategically harnessed to augment model certainty.
Open AccessAnalog In-Memory Computing with Uncertainty Quantification for Efficient Edge-based Medical Imaging Segmentation
This work investigates the role of the emerging Analog In-memory computing (AIMC) paradigm in enabling Medical AI analysis and improving the certainty of these models at the edge. It contrasts AIMC's efficiency with traditional digital computing's limitations in power, speed, and scalability. Our comprehensive evaluation focuses on brain tumor analysis, spleen segmentation, and nuclei detection. The study highlights the superior robustness of isotropic architectures, which exhibit a minimal accuracy drop (0.04) in analog-aware training, compared to significant drops (up to 0.15) in pyramidal structures. Additionally, the paper emphasizes IMC's effective data pipelining, reducing latency and increasing throughput as well as the exploitation of inherent noise within AIMC, strategically harnessed to augment model certainty.
- Open Access
Using the IBM Analog In-Memory Hardware Acceleration Kit for Neural Network Training and Inference
Analog In-Memory Computing (AIMC) is a promising approach to reduce the latency and energy consumption of Deep Neural Network (DNN) inference and training. However, the noisy and non-linear device characteristics, and the non-ideal peripheral circuitry in AIMC chips, require adapting DNNs to be deployed on such hardware to achieve equivalent accuracy to digital computing. In this tutorial, we provide a deep dive into how such adaptations can be achieved and evaluated using the recently released IBM Analog Hardware Acceleration Kit (AIHWKit), freely available at https://github.com/IBM/aihwkit. The AIHWKit is a Python library that simulates inference and training of DNNs using AIMC. We present an in-depth description of the AIHWKit design, functionality, and best practices to properly perform inference and training. We also present an overview of the Analog AI Cloud Composer, a platform that provides the benefits of using the AIHWKit simulation in a fully managed cloud setting along with physical AIMC hardware access, freely available at https://aihw-composer.draco.res.ibm.com. Finally, we show examples on how users can expand and customize AIHWKit for their own needs. This tutorial is accompanied by comprehensive Jupyter Notebook code examples that can be run using AIHWKit, which can be downloaded from https://github.com/IBM/aihwkit/tree/master/notebooks/tutorial.
Open AccessUsing the IBM Analog In-Memory Hardware Acceleration Kit for Neural Network Training and Inference
Analog In-Memory Computing (AIMC) is a promising approach to reduce the latency and energy consumption of Deep Neural Network (DNN) inference and training. However, the noisy and non-linear device characteristics, and the non-ideal peripheral circuitry in AIMC chips, require adapting DNNs to be deployed on such hardware to achieve equivalent accuracy to digital computing. In this tutorial, we provide a deep dive into how such adaptations can be achieved and evaluated using the recently released IBM Analog Hardware Acceleration Kit (AIHWKit), freely available at https://github.com/IBM/aihwkit. The AIHWKit is a Python library that simulates inference and training of DNNs using AIMC. We present an in-depth description of the AIHWKit design, functionality, and best practices to properly perform inference and training. We also present an overview of the Analog AI Cloud Composer, a platform that provides the benefits of using the AIHWKit simulation in a fully managed cloud setting along with physical AIMC hardware access, freely available at https://aihw-composer.draco.res.ibm.com. Finally, we show examples on how users can expand and customize AIHWKit for their own needs. This tutorial is accompanied by comprehensive Jupyter Notebook code examples that can be run using AIHWKit, which can be downloaded from https://github.com/IBM/aihwkit/tree/master/notebooks/tutorial.
- Open Access
- 28 November 2023
- Published by AIP Publishing in APL Machine Learning
- Vol. 1 (4)
Using the IBM analog in-memory hardware acceleration kit for neural network training and inference
Analog In-Memory Computing (AIMC) is a promising approach to reduce the latency and energy consumption of Deep Neural Network (DNN) inference and training. However, the noisy and non-linear device characteristics and the non-ideal peripheral circuitry in AIMC chips require adapting DNNs to be deployed on such hardware to achieve equivalent accuracy to digital computing. In this Tutorial, we provide a deep dive into how such adaptations can be achieved and evaluated using the recently released IBM Analog Hardware Acceleration Kit (AIHWKit), freely available at https://github.com/IBM/aihwkit. AIHWKit is a Python library that simulates inference and training of DNNs using AIMC. We present an in-depth description of the AIHWKit design, functionality, and best practices to properly perform inference and training. We also present an overview of the Analog AI Cloud Composer, a platform that provides the benefits of using the AIHWKit simulation in a fully managed cloud setting along with physical AIMC hardware access, freely available at https://aihw-composer.draco.res.ibm.com. Finally, we show examples of how users can expand and customize AIHWKit for their own needs. This Tutorial is accompanied by comprehensive Jupyter Notebook code examples that can be run using AIHWKit, which can be downloaded from https://github.com/IBM/aihwkit/tree/master/notebooks/tutorial.
- 28 November 2023
- Published by AIP Publishing in APL Machine Learning
- Vol. 1 (4)
Open AccessUsing the IBM analog in-memory hardware acceleration kit for neural network training and inference
Analog In-Memory Computing (AIMC) is a promising approach to reduce the latency and energy consumption of Deep Neural Network (DNN) inference and training. However, the noisy and non-linear device characteristics and the non-ideal peripheral circuitry in AIMC chips require adapting DNNs to be deployed on such hardware to achieve equivalent accuracy to digital computing. In this Tutorial, we provide a deep dive into how such adaptations can be achieved and evaluated using the recently released IBM Analog Hardware Acceleration Kit (AIHWKit), freely available at https://github.com/IBM/aihwkit. AIHWKit is a Python library that simulates inference and training of DNNs using AIMC. We present an in-depth description of the AIHWKit design, functionality, and best practices to properly perform inference and training. We also present an overview of the Analog AI Cloud Composer, a platform that provides the benefits of using the AIHWKit simulation in a fully managed cloud setting along with physical AIMC hardware access, freely available at https://aihw-composer.draco.res.ibm.com. Finally, we show examples of how users can expand and customize AIHWKit for their own needs. This Tutorial is accompanied by comprehensive Jupyter Notebook code examples that can be run using AIHWKit, which can be downloaded from https://github.com/IBM/aihwkit/tree/master/notebooks/tutorial.
- Open Access
- 28 September 2023
- Published by IOS Press
- p. 239-246
Pareto Rank-Preserving Supernetwork for Hardware-Aware Neural Architecture Search
In neural architecture search (NAS), training every sampled architecture is very time-consuming and should be avoided. Weight-sharing is a promising solution to speed up the evaluation process. However, training the supernetwork incurs many discrepancies between the actual ranking and the predicted one. Additionally, efficient deep-learning engineering processes require incorporating realistic hardware-performance metrics into the NAS evaluation process, also known as hardware-aware NAS (HW-NAS). In HW-NAS, estimating task-specific performance and hardware efficiency are both required. This paper proposes a supernetwork training methodology that preserves the Pareto ranking between its different subnetworks resulting in more efficient and accurate neural networks for a variety of hardware platforms. The results show a 97% near Pareto front approximation in less than 2 GPU days of search, which provides 2x speed up compared to state-of-the-art methods. We validate our methodology on NAS-Bench-201, DARTS, and Im-ageNet. Our optimal model achieves 77.2% accuracy (+1.7% compared to baseline) with an inference time of 3.68ms on Edge GPU for ImageNet, which yields a 2.3x speedup. Training implementation can be found:
- 28 September 2023
- Published by IOS Press
- p. 239-246
Open AccessPareto Rank-Preserving Supernetwork for Hardware-Aware Neural Architecture Search
In neural architecture search (NAS), training every sampled architecture is very time-consuming and should be avoided. Weight-sharing is a promising solution to speed up the evaluation process. However, training the supernetwork incurs many discrepancies between the actual ranking and the predicted one. Additionally, efficient deep-learning engineering processes require incorporating realistic hardware-performance metrics into the NAS evaluation process, also known as hardware-aware NAS (HW-NAS). In HW-NAS, estimating task-specific performance and hardware efficiency are both required. This paper proposes a supernetwork training methodology that preserves the Pareto ranking between its different subnetworks resulting in more efficient and accurate neural networks for a variety of hardware platforms. The results show a 97% near Pareto front approximation in less than 2 GPU days of search, which provides 2x speed up compared to state-of-the-art methods. We validate our methodology on NAS-Bench-201, DARTS, and Im-ageNet. Our optimal model achieves 77.2% accuracy (+1.7% compared to baseline) with an inference time of 3.68ms on Edge GPU for ImageNet, which yields a 2.3x speedup. Training implementation can be found: