Data Analyst Assistant
Resume:
Shahab Shams
***********@*****.***
******.*****@********.***
LinkedIn GitHub
RESEARCH INTEREST
I am a Machine Learning researcher with a particular interest in Graphical Models. I am currently work- ing on Approximate Inference and Algebraic Combinatorics. I have also been involved and interested in projects related to
• Spectral Theory
• Deep Learning
• Natural Language Processing
• Reinforcement Learning
EDUCATION
• Doctor of Philosophy Computer Science Aug. 2017 – Present University of Texas at Dallas Richardson, Texas
Advisor: Nicholas Ruozzi
• Master of Science Computer Science Sept. 2013 – Jan. 2016 Sharif University of Technology Tehran, Iran
Thesis: Extracellular Communication: Stochastic Modeling for Integrin Clustering and Activation. Advisor: Seyed Abolfazl Motahari
• Bachelor of Science Computer Science Sept. 2008 – Sept. 2013 Shiraz University Shiraz, Iran
Project: Optimizing Design of Steel Frame Structures by Utilizing Evolutionary Algorithms. PUBLICATIONS
• Counting Homomorphisms in Bipartite Graphs IEEE ISIT, July 2019. Shahab Shams, Nicholas Ruozzi, and Peter Csikvari
•
Evaluating the Impact of Program Features on Static Analysis Design Tradeoffs: A Java Numeric Analysis Case Study
Maliha Sarwat, Shahab Shams, Nicholas Ruozzi, Shiyi Wei Submitted to Software: Practice and Experience
• Markov Random Fields, Homomorphism Counting, and Sidorenko’s Conjecture On arXiv very soon
WORK EXPERIENCE
• Data Analyst Feb. 2016 - Jun. 2017
Center of Traffic and Transportation Research, Shiraz University Shiraz, Iran
• Teaching Assistant Jan. 2019 - Present
Probabilistic Graphical Models University of Texas at Dallas Numerical Methods for Machine Learning and AI
Machine Learning
PROGRAMMING SKILLS
Python
Tensorflow 2, NumPy, SciPy, Matplotlib, Pandas, scikit-learn MATLAB
C/C++
PROJECTS
• Weighted Schedule Sampling
For training an RNN-based generative model, one of the main unsolved problems is how to define the penalty function. A hybrid method of Scheduled Professor Forcing is commonly utilized to lessen the shortcomings of the existing methods. Modifying Schedule Sampling by assembling the feedback as a weighted sum of the network’s synthetic data and the real prefix does not resolve the issue in theory, but there are cases where it accelerates the convergence.
• Approximating the Partition Function for Globally Normalized Transition-Based NNs Andor et.al. introduced a Directed Neural Network for part-of-speech tagging and dependency parsing, in which beam inference is hired to approximate the partition function. One may think replacing beam inference with the Bethe Approximation enhances the efficiency. Although Bethe Approximation is a more complex method, it does not result in higher precision. It is because, in these particular tasks of Natural Language Processing, there are too few plausible choices with high probabilities that the simplistic beam search gives a good approximation of the partition function.
• Multiscale Graph Wavelet Transform
The Graph Convolutional Network is based on Spectral Graph Fourier transform gone through a series of severe simplifications. These simplifications are necessary to avoid overfitting, but they discard many advantages of using Spectral Fourier transform. Hiring Spectral Graph Wavelet transform helps to eliminate some of those simplifications because it modifies GCN so that it takes a mother wavelet with multiple scales as learnable parameters. However, the high number of pa- rameters is out of our control and the result is an extremely fast convergence into an uncompetitive model.
HONORS AND AWARDS
• National Competitive Examination for Post Graduate University Entrance Program(s) Feb 2013 Rank 37
• Iranian Mobile Phone Innovation Contest, Generic Mobile Software Development Mar 2012 2nd place
• Iranian Mobile Phone Innovation Contest, Nokia Software Software Development Mar 2012 3rd place
• Asian ACM-ICPC Contest Dec 2011
8th place
Contact this candidate