M. H. Ismail

Academic

Career

Verbatim text

MorphNet – Algebraic Methods in Graph Neural Network Design
We have developed a novel approach to learning over non-Euclidean data (primarily represented as graphs) using an algebraic lens along with physical models.
MorphNet works by viewing local subgraphs as instances of the n-body problem from astronomy and uses a discrete approximation of the Hamiltonian to train the network to find the optimal mass configuration and choreography. Afterwards, unidentified vertices may have their choreography and mass realized – by viewing masses as indicators of class, we attain a classification.
To demonstrate our work, we have created a recommender system as a web application. Acting on a movie dataset, users may select preferences – taking advantage of the manifest topology, MorphNet uses the graph-of-edges to make recommendations.
In addition, we have developed a simulation to visualize the behavior of graph neural network mechanisms – essentially, the influence of graph features (e.g. vertex features, edge features, etc.) are propogated through and demonstrated on the simulation in real-time. This allows one to contrast different graph deep learning mechanisms.
Group Member
i190416 - M. H. Ismail
Supervisor
Dr. Omer Beg
Co-Supervisor
Mr. Irfan Ullah (Humanities)
Technology Used:
Python, PyTorch, Flask, React.JS, D3.JS
Supervisor Name:
Dr. Mirza Omer Beg
Co-Supervisor Name:
Mr. Irfan Ullah (Humanities Dept.)
Group Members:
M. H. Ismail (i19 - 0416)
Abuzar (i19 - 0531)
Abdullah Bilal (i15 - 0616)
Tools & Technologies
Python, D3, Flask, GitHub, React, Docker

AI enrichment