Hanady GEBRAN

Pooling techniques in hybrid quantum-classical convolutional neural networks

This work proposes hybrid quantum-classical convolutional neural networks (QCCNNs) for medical imaging classification. The study focuses on exploring different pooling techniques in QCCNNs for 2D medical image classification tasks. By simulating the network using the pennylane framework, comparable or improved performance is observed compared to equivalent classical models and QCCNNs without pooling. These findings highlight the need for further investigation into architectural choices in QCCNNs for future applications.

You can access the letter of reference from Fraunhofer here. , the full master thesis report here. and the accepted paper here.

Pennylane Qiskit Quantum computing

Cognition-based clustering and its use in relevant descriptions

This work aimed to consider clustering as a cognitive process and to exploit human reasoning to achieve it. The ultimate goal was to achieve optimal clustering with few examples and no parameters. To this end, I have relied on heuristic approaches beyond K-means, based on Kolmogorov complexity (which measures the intuitiveness of a feature the intuitiveness of a piece of information) and contrast (which compares an object to a given prototype), with the aim of mimicking human decision-making considerations.

You can access the full report here.

Python Pandas Scikit-learn

Formalizing mathematical statements in LEAN

FAIR (Facebook AI Research) is in the midst of a race with Google and INRIA to determine the extent to which formal languages (such as LEAN) can be used to produce and prove mathematical statements. FAIR has therefore contacted leading grandes écoles to create mathematical datasets in LEAN.

I was part of the team at École Polytechnique and worked on International Mathematical Olympiad problems.

Solving the robotic reaching task using a spiking neural network

The goal of this project was to solve a reaching task in the Neurorobotics platform using a manipulator robot and spiking neural networks (SNNs).

In each episode, a target point was randomly generated and the goal was to control the angles of the manipulator joints so that the end effector would reach that point. The observations were simulation ground truth data such as target pose, current joint angles, and velocities.

This was a joint project: we each tailored an RL algorithm to the SNN with an appropriate design to encode and decode the inputs and outputs: I chose TD3 and integrated it, then my partner coded A2C.

The deployment was as follows: RL container : Training code + GRPC client; Backend containers : NRP backend (Gazebo, NRP, ROS) + GRPC server; Frontend container : REST API communication with NRP Backend + NRP UI on a localhost port (HTTP).

You can see the full report here.

Python Docker Gazebo NRP LRZ cloud Git

StreamifAI

The objective of this project was to build a recommendation system in less than 36 hours for HackaTUM as part of the Hubert Burda Media challenge: we won second place against 20 other teams.

The goal was to avoid having the client grow tired of making decisions about which movie to watch next. The concept we proposed was: take a selfie with your family or a group of friends and a stream will be attributed to you taking into account the compromise factor in an innovative way, using mathematical constraints.

I proposed the compromAIse paradigm and implemented it, cleaned the data and built the recommendation system itself.

You can find more information about the project here. and more information about the hackathon here. .

Energy System Design in Europe: A Multi-Objective Approach

The objective of this project was to develop a comprehensive method for modeling renewable energy systems in order to achieve renewable energy production that meets all the energy needs of selected European countries.

The aim was to be realistic from the point of view of budget optimization, but also to avoid public opposition. Moreover, in order to examine the interactions with a third objective dimension, another objective was the question of employment, i.e. to generate a maximum of jobs in the production of green energy.

We worked in pairs on this project; my main task was to collect data from several official European sites and to choose the objective functions and code them on top of a simple apriori model whereas my partner was mostly coding the other methods once we had a satisfactory model.

You can see the full report here.

You can download the zip file of the code submitted for this project here.

Ampl

Plants vs. Zombies: Reinforcement learning to a tower defense game

In a four-person team, we developed several reinforcement learning methods (policy gradient with or without actor critique module, Deep Q-Network and Double Deep Q-Network) and successfully trained agents to play the plant vs. zombie game we developed in a game engine, linked to an OpenAI gym environment. We presented the superiority of policy-free methods for this environment and how hyperparameters and observation features can be manipulated to improve agent performance.

I worked mostly on the environment and the balance of the game.

You can see the full report here.

You can see the git repository submitted for this project here.

Python Pytorch Git

Software project: delivery system

The project aimed to build a delivery system with the following specifications:

Back-end: Spring, MongoDB, Java.

Front-end: React, Javascript.

Embedded system: Python in a Raspberry Pi

Deployment: Docker.

I was in charge of the embedded system (solo) and the backend (team of 3), and I had a secondary role with the deployment. I was also in charge of presenting the project and doing a complete code walkthrough in front of the jury.

You can download the zip file of the code submitted for this project here.

Infinitely wide Transformer networks and their Laplace operators

In collaboration with the TUM Mathematics Department and the Chair of Scientific Computing, I studied the relationship between the integral operator of kernels associated with an infinite-width network and the Laplace-Beltrami operator on the manifolds where the data lies; in addition to a small application which was a transformers on an IMDb dataset.

You can see the full report here.

You can download the ipynb file of the code submitted for this project here.

Python Jax