Responsible: Gianni Franchi (ENSTA Paris / U2IS)
Teaching Assistants: Project tutors and PhD mentors depending on the topic
Assessment: Code quality (40%), Written report (30%), Final presentation (30%)
Course Description
The CSC_5IA21 – Project IA course is a group-based project designed to immerse students in real-world artificial intelligence challenges. Students form teams of 4 to 6 members and select one of the proposed research-oriented projects. Each project focuses on applying and extending AI methods across domains such as deep learning, explainability, generative models, multimodal fusion, and trustworthy AI.
The objective of this course is to help students develop research autonomy, teamwork skills, and an applied understanding of modern AI methods while working under the supervision of expert tutors. The project outcome includes code implementation, a written report, and a final oral presentation.
Available Projects (2025–2026)
P1 — Deep Fakes Image Detection: Develop and benchmark deep neural models capable of detecting manipulated or generated images from real ones.
P2 — Deep Fakes Sound & Multimodal Generation: Explore multimodal deepfake generation and detection by combining audio and video modalities.
P3 — Sound Unmixing: Implement and evaluate machine learning methods for separating overlapping sound sources. (Tutor: Mathieu F.)
P4 — Explainable AI & Mechanistic Interpretability: Analyze internal mechanisms of neural networks to improve interpretability and understanding of learned representations.
P5 — Physics-AI and Uncertainty: Combine physics-based modeling and machine learning, focusing on uncertainty estimation and trustworthiness of predictions.
P6 — Semantic Segmentation in Open-World Datasets: Train segmentation models on complex real-world datasets and address domain shift and OOD generalization.
P7 — LLMs and Hallucination: Study hallucination behaviors in Large Language Models and evaluate statistical or structural mitigation strategies.
P8 — Synthetic Data Generation for RAG Evaluation: Build synthetic datasets to evaluate Retrieval-Augmented Generation (RAG) systems and analyze their robustness.
P9 — Anomaly Detection in Microscopic Images: Develop image anomaly detection techniques for defect identification in microscopic data.
P10 — Analysis of Deep Learning Architectures and Training Dynamics: Examine how architecture choices and optimization parameters affect model behavior and performance.
P11 — Embodied AI, LVLMs & Vision-Language-Action Models: Explore the integration of large multimodal models in embodied robotics and action understanding tasks.
P12 — Efficient Image Generation: Investigate advanced optimization techniques such as LoRA, pruning, quantization, and ControlNet for faster and energy-efficient image generation.
Schedule Breakdown
- October to December: Conduct 3–4 sessions with your tutors (initial planning, goal setting, and early progress checks).
- January to March: Conduct 6–7 sessions focused on implementation, debugging, and refinement.
Total: 10–11 sessions across the semester. Schedule meetings in advance to ensure tutor availability and consistency. Each group must document progress and decisions after every session using the Discord project channel, following the template shared in class.
Key Notes
- Attendance is mandatory for all scheduled sessions.
- Missing more than 3 sessions will result in a −1 point penalty on the final grade.
- Students are expected to collaborate effectively within their groups and communicate regularly with their tutors.
- All results, code updates, and feedback summaries must be shared on the Discord project space.
Evaluation Criteria
- Code (40%) — Implementation quality, documentation, efficiency, and originality.
- Written Report (30%) — Clarity of objectives, methodology, results, and analysis.
- Final Presentation (30%) — Oral defense, clarity, teamwork, and ability to discuss the scientific and technical contributions.