1. Smart Assist System for Blind people 
2. Federated Learning Based Proactive Content Caching in Edge Computing
This work is done by collaboration between LA Rochelle and AUC Universities. In this work, We build neural collaborative filtering technique to tackle proactive content caching. 3. End-To-End self driving cars using Imitation learning with future prediction using GANs 
In this work, I will tackle self driving car problems following the End to End approach, by building a deep learning agent which will be able to control the vehicle by outputting throttle, steer and brake value. It was proven in several occurrence that taking historical information will be a huge added value to the agent robustness. Historical information may be the direct previous sensor readings or the distribution of the network outputs. There are a lot of ways to achieve that for example by using Recurrent neural networks or even by simply stacking the previous frames from the sensor readings. What if taking the advantage of the future too, using video prediction techniques, we could predict the future frames on the basis of previous frames. Then stacking them with the current and the previous frames to increase our model robustness. 4. End-To-End Deep Path Planning and Automatic Emergency Braking Camera Cocoon-based Solution
Deep Learning is contributing greatly in many automotive applications like: autonomous driving, and augmented reality. Fully autonomous driving is one of hot research fields nowadays because of its difficulty. Path planning and the automatic emergency braking are considered as two of the main functionalities achieving fully autonomous driving vehicle. Instead of difficult prerequisite of traditional path planning, deep path planning is proposed depending on camera cocoon installed covering 360 degrees around the vehicle. Due to the absence of benchmarks serving our idea, we build our own benchmark based on CARLA Simulator covering difficult and various scenarios or situations that the vehicle may expose. Our proposed model is robust due to its high generalization capability in differentiating between lane over taking and stop through braking. 5. Teach a quadcopter how to fly using RL
6. Categorize animals in the wild using object detection
