Channel-Remix and CarveVFM ↗
Manuscripts on task-specific channel recombination and model compression are under review.
CarveVFM on arXiv ↗ MBZUAI · ABU DHABILotfi Abdelkrim
Lotfi Abdelkrim
Mecharbat
PhD student in Machine Learning
Foundation ModelsModel CompressionEfficient Deep LearningModel AdaptationKnowledge DistillationMeta-learningMedical ImagingAutoMLNeural Architecture Search
I’m a PhD student at Mohamed bin Zayed University of Artificial Intelligence, supervised by Prof. Mohammad Yaqub and Prof. Martin Takáč.
I work on adapting and compressing foundation models for new tasks with limited data and compute. My interests include efficient deep learning, knowledge distillation, and meta-learning, as well as AutoML and neural architecture search.
Arabic · English · French
MY PHD / A RESEARCH MAPAutomatic specialization
Automatic specialization
of foundation models.
My PhD studies how to adapt and compress a pretrained foundation model for a task, given calibration data and deployment constraints.
THESIS FRAMEWORK Proposed joint system
Task calibration data
DImage encoder
Text sample
Text encoder“a small bird”↓ Task features D
Hardware & budget
cms LatencyMB MemoryJ Energy
01 / MAIN INPUTM₀
Pretrained
foundation model
Pretrained representations
Full model depth and width
02 / META-LEARNING
Specialization
controller
Learn task–configuration
relationships across datasets
α* adaptationφ₀* initializationp* compression
03 / SPECIALIZE
α*, φ₀*Adaptation
Placement · rank · initialization
Pretrained channel recombination
Compression
Depth · width · structured masks
Capacity allocation and recovery
04 / OUTPUTM*
Task-specific
model
Task performanceDeployment constraints
Pretrained weights Adaptation Retained capacityIllustrative model structures
A pretrained model, task images, and hardware budgets define the specialization problem.
PAPERS / SELECT A REFERENCE
The joint controller and extension beyond vision are proposed work. References show each method’s contribution; MaskVFM is ongoing.
NEWS
All updates ↗Research updates.
Winning with the university football team
I captained the MBZUAI team to a tournament win.
CARDPOSTS
View Cardposts ↗Cardposts
No Cardposts published yet.
01 / RESEARCH INTERESTS
Research focus areas.
I work on model design, adaptation, and deployment, mainly for vision tasks.
Foundation models & adaptation
Adapting pretrained models to new tasks with meta-learning, adapters, and LoRA.
Model compression & efficient deep learning
Making models smaller and cheaper to run through pruning, channel recombination, and hardware-aware deployment.
Knowledge distillation & meta-learning
Transferring knowledge from pretrained models and previous tasks to smaller task-specific models.
AutoML & neural architecture search
Automating architecture and training choices under task and hardware constraints.
SELECTED WORK
Browse all research ↗Selected research
ICLR 2027 · under review
Channel-Remix
Build compact task-specific models by selecting and sparsely recombining pretrained channels instead of relearning new weight directions.
View project ↗AAAI 2027 · under review
CarveVFM
Jointly search capacity allocation, removal operators, and recovery to extract a compact dense model.
View project ↗MICCAI 2026
MetaVFM
Use meta-learning to select and initialize adapters for a new medical imaging task.
View project ↗MICCAI 2025
MedNNS
Embed medical datasets and supernet subnetworks in a shared space, so a new task can retrieve a matching architecture and initialization.
View project ↗EDUCATION & EXPERIENCEWhere I’ve studied
Education & experience ↗Where I’ve studied
and worked
ABOUT ME
Background and interests
I studied computer science in Algiers, worked on railway vision systems in France, and moved to Abu Dhabi for research at NYUAD and then MBZUAI.
I like football, trekking, and history. I’ve also worked on player and ball tracking for amateur football videos.
About me ↗CONTACT
Contact
You can email me about my research or a possible collaboration.