Bilal FAYE
PhD in Artificial Intelligence | Researcher | Educator | AI Vision & NLP Specialist.
📍 Saint Denis, France | ✉️ biljolefa@gmail.com | LinkedIn
PhD in Artificial Intelligence | Researcher | Educator | AI Vision & NLP Specialist.
📍 Saint Denis, France | ✉️ biljolefa@gmail.com | LinkedIn
Diffusion models are a class of generative models that iteratively denoise data from random noise to generate samples. They have demonstrated remarkable performance in image, audio, and even text generation.
Popular models include DDPM, Stable Diffusion, and Imagen. Despite their success, they are computationally expensive and require long inference times.
Our Contribution: We propose a lightweight conditioning-aware sampling strategy that reduces inference steps by learning adaptive guidance schedules. This technique bridges control and efficiency, enabling faster and controllable generation with fewer resources—opening opportunities for real-time applications.
Reinforcement learning is widely used to fine-tune LLMs to better align with human preferences. RLHF (Reinforcement Learning with Human Feedback) combines supervised fine-tuning, preference models, and Proximal Policy Optimization (PPO).
However, RLHF is limited by instability and reliance on carefully crafted preference datasets. DPO (Direct Preference Optimization) simplifies the process using prompt-preferred-rejected triplets but depends heavily on quality data, which is hard to collect.
DRO (Direct Reward Optimization) goes further by using (prompt, answer, reward) tuples to jointly learn the policy and value functions, which remains a challenging optimization problem.
Our Contribution: We develop a hybrid DRO framework that decouples policy and value learning using modular actor-critic heads with memory-based priors. This structure reduces learning instability and enhances scalability across tasks with noisy reward functions.
Challenge: Most multimodal LLMs are expensive and hard to scale.
Objective: OmniLLM provides an efficient framework to extend any LLM to multimodal tasks with minimal cost.
Challenge: Many African languages are low-resource and underrepresented in NLP.
Objective: Build efficient multilingual LLMs for African languages using transfer learning and low-resource techniques.
Challenge: Open-vocabulary detection is expensive and data-hungry.
Solution: LightMDETR is a modular framework reducing training cost while maintaining performance. Variants like LightMDETR-PLUS push the efficiency frontier.
Challenge: High training cost and reliance on large paired datasets.
Solution: OneEncoder aligns multiple modalities with a unified encoder, supporting zero-shot and domain-specific tasks.
Challenge: Traditional normalization techniques assume simple distributions or use expensive clustering.
Solution: Context Normalization introduces supervised and unsupervised normalization using prior knowledge without EM. Improves efficiency and generalization.