Institution profile

Université Mohammed Premier Oujda

Academic institutionafrica · ma
Official website
Research library12linked papers
Opportunities0open roles
Selected work

Representative Papers

Motion-Compensated Weight Compression

May 23, 2026

This work addresses a critical limitation in existing neural network compression methods: their neglect of cross-layer redundancy arising from functional symmetries, such as permutation invariance among hidden units and attention heads. To overcome this, the authors propose a novel framework that aligns symmetric blocks across layers via motion compensation, transforming weight sequences into predictable structures. The approach introduces a lightweight per-layer predictor, a rate–distortion-optimized entropy model, and a keyframe scheduling mechanism to efficiently encode quantized residuals. During decoding, inverse alignment enables rapid weight reconstruction. By explicitly modeling cross-layer alignment for the first time, the method achieves substantial improvements over state-of-the-art quantization and learned compression techniques on Transformer-based language modeling and vision classification tasks, significantly advancing the rate–accuracy Pareto frontier while preserving inference speed.

0 citationsRead paper

When Bits Break Recourse: Counterfactual-Faithful Quantization

May 16, 2026

This work addresses the tension between low-bit quantization and algorithmic recourse, showing that aggressive quantization—while preserving predictive accuracy—can severely undermine counterfactual explainability by invalidating or drastically inflating the cost of recourse. The paper formally characterizes this trade-off through three criteria for counterfactual sensitivity: validity, cost, and directional stability, and introduces two novel metrics, Validity Drop (VD) and Counterfactual Recourse Gap (CRG), to quantify recourse degradation. To reconcile efficiency with explainability, the authors propose Counterfactually Faithful Quantization (CFQ), a method that jointly optimizes prediction accuracy and recourse fidelity under a global bit budget via teacher-guided target constraints, mixed-precision bit allocation, and boundary perturbation theory. Experiments on Adult, German Credit, and COMPAS datasets demonstrate that CFQ significantly outperforms baselines, achieving comparable accuracy while markedly improving both VD and CRG.

0 citationsRead paper

Drift-to-Action Controllers: Budgeted Interventions with Online Risk Certificates

Mar 09, 2026

This work addresses the lack of safe and controllable response mechanisms in machine learning systems under distribution shift by proposing a monitoring framework grounded in constrained safe decision-making. The approach identifies shift types at the perception layer and, for the first time, introduces an online risk certificate mechanism that provides a valid, anytime upper bound on current risk, thereby triggering tiered intervention strategies. Key technical components include unlabeled signal modeling, delayed label querying, test-time adaptation, and active abstention. Experiments on WILDS Camelyon17, DomainNet, and synthetic data streams demonstrate that the method achieves near-zero safety violations and rapid recovery, significantly outperforming existing baselines.

0 citationsRead paper
Recent publications

Latest Papers

Motion-Compensated Weight Compression

May 23, 2026

This work addresses a critical limitation in existing neural network compression methods: their neglect of cross-layer redundancy arising from functional symmetries, such as permutation invariance among hidden units and attention heads. To overcome this, the authors propose a novel framework that aligns symmetric blocks across layers via motion compensation, transforming weight sequences into predictable structures. The approach introduces a lightweight per-layer predictor, a rate–distortion-optimized entropy model, and a keyframe scheduling mechanism to efficiently encode quantized residuals. During decoding, inverse alignment enables rapid weight reconstruction. By explicitly modeling cross-layer alignment for the first time, the method achieves substantial improvements over state-of-the-art quantization and learned compression techniques on Transformer-based language modeling and vision classification tasks, significantly advancing the rate–accuracy Pareto frontier while preserving inference speed.

0 citationsRead paper

When Bits Break Recourse: Counterfactual-Faithful Quantization

May 16, 2026

This work addresses the tension between low-bit quantization and algorithmic recourse, showing that aggressive quantization—while preserving predictive accuracy—can severely undermine counterfactual explainability by invalidating or drastically inflating the cost of recourse. The paper formally characterizes this trade-off through three criteria for counterfactual sensitivity: validity, cost, and directional stability, and introduces two novel metrics, Validity Drop (VD) and Counterfactual Recourse Gap (CRG), to quantify recourse degradation. To reconcile efficiency with explainability, the authors propose Counterfactually Faithful Quantization (CFQ), a method that jointly optimizes prediction accuracy and recourse fidelity under a global bit budget via teacher-guided target constraints, mixed-precision bit allocation, and boundary perturbation theory. Experiments on Adult, German Credit, and COMPAS datasets demonstrate that CFQ significantly outperforms baselines, achieving comparable accuracy while markedly improving both VD and CRG.

0 citationsRead paper

Drift-to-Action Controllers: Budgeted Interventions with Online Risk Certificates

Mar 09, 2026

This work addresses the lack of safe and controllable response mechanisms in machine learning systems under distribution shift by proposing a monitoring framework grounded in constrained safe decision-making. The approach identifies shift types at the perception layer and, for the first time, introduces an online risk certificate mechanism that provides a valid, anytime upper bound on current risk, thereby triggering tiered intervention strategies. Key technical components include unlabeled signal modeling, delayed label querying, test-time adaptation, and active abstention. Experiments on WILDS Camelyon17, DomainNet, and synthetic data streams demonstrate that the method achieves near-zero safety violations and rapid recovery, significantly outperforming existing baselines.

0 citationsRead paper