Ph.D. Computer Science
2018 - 2024
- Advisor: Dr. Xiaoli Fern
- Graph Neural Networks (GNNs) for molecule generation
razaa@oregonstate.edu
Phone
+541-360-9372
Location
Oregon, USA
About me
Who/What am I?
Hello, I'm Ali Raza (he/him)
I am a research scientist [machine learning] at Meta (Facebook/Instagram).
Fei-Fei Li, Stanford Artificial Intelligence Lab Director, said at WIRED25 Summit:
"What’s really important is putting humanity at the center."
Be it the first-ever optical character recognition (OCR) system for Urdu (the national language of Pakistan, perso-arabic nastalique writing styles) language or a distributed GIS-based real-time syndromic surveillance system during the dengue epidemic in Pakistan; I always invested myself in a project/research that is beneficial for human society, natural environment or the synergy between them.
I am a tech enthusiast. I spend my time learning about new technology.
I am a hobby photographer who likes street photography. I play badminton and cricket. I am learning how to play guitar. I am a really good singer. People who have heard me may disagree and they are probably right.
My Education
M.Sc. Computer Engineering
2013 - 2015
- Advisor: Dr. Sadiq M. Sait
- Thesis: Energy-aware Resource Allocation Heuristics for Efficient Management of Data Centers
B.Sc. Electrical Engineering
2008 - 2012
- Advisors: Dr. Haroon Atique Babri & Dr. Naeem Ayyaz
- Thesis: A Distributed GIS-based Real-Time Syndromic Surveillance System
My Expericence
Research Scientist (Machine Learning)
2024 - Present• Working with the IFR (In‑Feed recommendations) team
Graduate Research Assistant
2018 - PresentUsing machine learning and deep learning to design graph neural networks (GNNs) to process graph data, make predictions, and explain those predictions. Worked on research projects related to computer vision and recommendation systems
- Designed and trained an explainable message passing neural network (MPNN) with soft attention mechanism to predict (mean absolute error - MAE: 0.0616) simulated CO2 adsorption in MOFs and provide insights into what substructures of the MOFs are important for the prediction
- Implemented an MPNN to predict the atomic partial charges on MOFs under a hard charge-neutral constraint obtaining an MAE of 0.025 as compared to 0.154 of the previous work
- Designed a framework utilizing a set of predefined transformations and a global aggregate layer to encode periodic images to the latent space. Our approach resulted in mean square error - MSE: 0.0478 as compared to 0.0585 of the stacked autoencoder
- Implemented a recommendation system to predict missing adsorption properties of nanoporous materials and produce a map of COFs (covalent-organic frameworks), wherein COFs with similar (dissimilar) adsorption properties congregate (separate)
- Developed an LSTM based multi-task learning model for predicting blood glucose levels in patients with diabetes
Software Engineer PhD Intern
Jun 2022, Sept 2023Worked with the IFR (In-Feed recommendations) team to build implicit interests based on user engagements to recommend unconnected content
- Generated group-group similarity graph using user-group engagements
- Used clustering algorithm to compute group-based implicit interest
- Used group-based implicit interest clusters to generate user and object representations
- Implemented user-interest-item generator for unconnected recommendations and run online experiments
Graduate Teaching Assistant
2018 - 2019
- CS 261 Data Structure
[Winter, Spring, Summer 2019] - CS 271 Computer Architecture & Assembly Language
[Fall 2018]
IT Support Specialist
2016 - 2018
- Network management and diagnosis
Lecturer
2016 - 2018
- Taught computer information technology courses
Graduate Teaching Assistant
2015 - 2015
- COE 241 Data and Computer Communication
[Fall 2015]
Graduate Research Assistant
2013 - 2015Worked on optimizing power utilization in data centers using non-deterministic heuristics
- Tested and analyzed the interaction between server consolidation and thermal management.
- Implemented Tabu Search, Simulated Annealing, and Genetic algorithms to optimize cooling and computational power.
- Designed a goodness function and engineered Simulated Evolution to reduce up-to 20% power utilization in data centers for offline VM placement problem
Associate Research Officer
2012 - 2013Worked with the Pre-processing team to develop an Optical Character Recognition (OCR) System using Java for Urdu (national language of Pakistan) Perso-Arabic Nastalique writing styles.
- Developed grammar-based modules for Mainbody-Diacritic Association and Ligature formation with an accuracy of 98.7%.
- Improved Connected-Component Disambiguation, from 80% to 98.87%, by implementing Diacritic Recognizer using j48 decision tree.
- Implemented modules for Figure, Pepper noise, and Page Frame detection with an accuracy of 96.5%
Adjunct Instructor
2012 - 2013
- Introduction to computing
[Spring 2013] - Applied Electricity
[Fall 2012, Spring 2013]
Undergraduate Researcher
2011 - 2012Worked with the C-SVAR team to develop a centralized disease surveillance system based on dis-tributed network of hospitals to address the infectious disease surveillance challenges in Pakistan
Research Intern
June 2011 - Aug 2011
- Developed an Application for Android platform using Java for displaying basic specs of the device and running Multicore Processor Architecture and Communication (MPAC) micro-benchmarks for Memory, Cache, Processor, and network so that users can compare the performance of different Android devices.
- Developed an application for iOS using C# that displays general information and specs.
- Used Scratch Box and Qemu for Cross Compilation and Emulation for multi-core ARM
- Performed as an assistant instructor in a short course Programming Multi-Core Processors.
Publications
" Non-equilibrium molecular geometries in graph neural networks."
Raza, A., Henle, A. ,&Fern, X.(2021)"Towards explainable message passing networks for CO2 adsorption in metal-organic frameworks."
Raza, A., Waqar, F., Sturluson, A., Simon, C., & Fern, X. Machine Learning for Molecules workshop at NeurIP (2021)"A recommendation system to predict missing adsorption properties of nanoporous materials."
Sturluson, A., Raza, A., McConachie, G. D., Siderius, D., Fern, X., & Simon, C. (2021)"Towards explainable message passing networks for CO2 adsorption in metal-organic frameworks."
Raza, A., Waqar, F., Sturluson, A., Simon, C., & Fern, X. Machine Learning for Molecules workshop at NeurIP (2021)"Message passing neural networks for partial charge assignment to metal–organic frameworks."
Raza, A., Sturluson, A., Simon, C. M., & Fern, X. Journal of Physical Chemistry C 124.35 (2020): 19070-19082"Engineering simulated evolution for integrated power optimization in data centers."
Sait, S. M., & Raza, A. Soft Computing 22.9 (2018): 3033-3048.Talks/Presentation
Machine Learning for Molecules workshop at NeurIPS Dec 2021
Machine Learning for Molecules workshop at NeurIPS Dec 2020
2020 Doctoral Consortium on Computational Sustainability, Oct 2020
7th Saudi Student Conference, 2016
Awards & Honours
- KFUPM Scholarship during two years of graduate study (2013-2015)
- Merit Scholarship by the University for excellent academic performance during the four years of undergraduate study (2008-2012)
- Gold Medal for scoring top position in the Higher Secondary School (2008)
- Shahbaz Shareef Laptop Award for maintaining high CGPA in the University (2012)
- Certificates of Appreciation for contributing in organizing R & D week (2009), Technofest (2010, and IETEC (2011) by The Institution of Engineering and Technology (IET), UET Chapter
- Certificate of Appreciation for contributing in organizing Conference on Language & Technology by Al-Khawarizmi Institute of Computer Science Lahore, Pakistan(2012)
- Certificate of Appreciation for working as an assistant instructor in a short course Programming Multi-Core Processors A Hands-On Approach for Embedded, Mobile, and Distributed Systems Development by Al-Khawarizmi Institute of Computer Science Lahore, Pakistan (2011)
Activities
- Member of Pakistan Student Association (PSA), OSU 2018 - Present
- Member of Pakistan Engineering Council (PEC) 2013 - Present
- Member of KFUPM Volunteer Group 2015 - 2018
- Served as a Volunteer in Pakistani Hajj Volunteer Group (PHVG), KSA 2014
- Human Resource (HR) Coordinator of IET UET Chapter 2011 - 2012
- Member of Management Team, IET UET Chapter 2009 - 2011
Selected Projects
Some of my highlighted projects.
#python, #pytorch, #machine learning
Towards explainable message passing networks for CO2 adsorption in metal organic frameworks
Metal-organic framework (MOFs) are nanoporous materials that could be used to capture carbon dioxide from the exhaust gas of fossil fuel power plants to mitigate climate change. In this work, we design and train a message passing neural network (MPNN) to predict simulated CO2 adsorption in MOFs. Towards providing insights into what substructures of the MOFs are important for the prediction...
Read More
#python, #pytorch, #machine learning
Message Passing Neural Networks for Partial Charge Assignment to Metal-Organic Frameworks
Virtual screenings can accelerate and reduce the cost of discovering metal-organic frameworks (MOFs) for their applications in gas storage, separation, and sensing. In molecular simulations of gas adsorption/diffusion in MOFs, the adsorbate-MOF electrostatic interaction is typically modeled by placing partial point charges on the atoms of the MOF. For the virtual screening of large libraries of MOFs...
Read More
#python, #pytorch, #machine learning
Latent representation of periodic images
Learning efficient latent space representation of the input space in an unsupervised manner has been a major research topic in machine learning. The latent representation can be used not only for dimensionality reduction by capturing the hidden structure of the data, but it can also be used to generate new data, either by interpolation or by sampling. Periodic images consist of infinite copies of their unit cells. A unit cell contains enough information to describe the whole image. Latent representations of these unit cells can correspond to the information-rich fingerprint of the whole image that encodes its salient features and can be used for features reduction, characteristic prediction, and synthesis of new images with given properties. However, a single periodic image can have a large number of unit cells and encoding them to a single point in the latent space is not trivial. In this paper, we propose a framework to encode periodic images while taking care of the boundary effect. We evaluate our approach with the vanilla autoencoder.
Read More
#python, #tensorflow, #machine learning
Multi Task Learning using LSTM
Predicting future blood glucose levels permits diabetes patients to take necessary action before imminent hyperglycemia and hypoglycemia. We used a deep learning network including long-short-term memory (LSTM) in multi‑task learning of blood glucose from time‑series data.
Read More
#python, #pytorch, #machine learning
Solving Sokoban Game
Artificial intelligence (AI) research has developed an extensive collection to methods to solve state-space problems. These techniques have been successful for a wide range of games like chess and checker. Sokoban is a simple puzzle game yet it has variety of problem instances with wide range of complexity from easy to extremely difficult. Furthermore, the powerful restriction of actions makes it an interesting benchmark for comparing different AI search algorithms. In this work, we evaluate six different search algorithms; Depth First, Breadth First, Uniform Cost, A*, and Monto Carlo search algorithms. Furthermore, we explore different heuristic functions. We use a wide range of problem instances for evaluation purposes.
Read More
#python, #pytorch, #machine learning
Automatic Environment Generation to Generalize Agents
Reinforcement learning(RL) techniques are a sub-field of machine learning approaches, in which an agent tries to learn the dynamics of an environment by trial and error by having interaction whithin the environment. The goal of the agent is to get the best feedback from the environment, which is defined as reward. Given the state of the environment, agent takes an action to maximize the expected future reward. The function that maps states to actions are called policies. However, it is difficult for the agents to learn a general policy that applies across similar environments. Furthermore, they do not get reasonable performances on the same environments of varying difficulty level. To address these problems, we introduce a new pipeline for generating environments with varying difficulty levels to improve the agents’ performances.
Read More
#python, #pytorch, #machine learning
Finding fundamental matrix, epipoles and epipolar lines
In this work, we used manual points and SIFT keypoints detector. more detail coming..
Read More
#java, #machine learning
Optical Character Recognition (OCR) System for Urdu (national language of Pakistan) Perso-Arabic Nastalique writing styles
Worked with the Pre-processing team to develop an Optical Character Recognition (OCR) System using Java for Urdu (national language of Pakistan) Perso-Arabic Nastalique writing styles.
Read More
#matlab, #heuristics, #combinatorial optimisation
Engineering simulated evolution for integrated power optimization in data centers
Cloud computing has evolved as the next generation platform for hosting applications ranging from engineering to sciences, and from social networking to media content delivery. The numerous data centers, employed to provide cloud services, consume large amounts of electrical power, both for their functioning and their cooling. Improving power efficiency, that is, decreasing the total power consumed, has become an increasingly important task for many data centers for reasons such as cost, infrastructural limits, and mitigating negative environmental impact. Power management is a challenging optimization problem due to the scale of modern data centers. Most published work focuses on power management in computing nodes and the cooling facility in an isolated manner. In this paper, we use a combination of server consolidation and thermal management to optimize the total power consumed by the computing nodes and the cooling facility. We describe the engineering of an evolutionary non-deterministic iterative heuristic known as simulated evolution to find the best location for each virtual machine (VM) in a data center based on computational power and data center heat recirculation model to optimize total power consumption. A “goodness” function which is related to the target objectives of the problem is defined. It guides the moves and helps traverse the search space using artificial intelligence. In the process of evolution, VMs with high goodness value have a smaller probability of getting perturbed, while those with lower goodness value may be reallocated via a compound move. Results are compared with those published in previous studies, and it is found that the proposed approach is efficient both in terms of solution quality and computational time.
Read More
#C, #machine learning
A Distributed GIS-based Real-Time Syndromic Surveillance
Each year, millions of Pakistanis are exposed to, and infected with, deadly pathogens including hepatitis, tuberculosis, malaria, and now dengue. Lack of a robust infrastructure for the timely collection, reporting, and analyses of Dengue Epidemic (DE) data undermines epidemic preparedness and poses serious health challenges to the general public in Pakistan. In fact, monitoring of the outbreak and response to any natural or man-made infectious disease (ID) is non-existent in the country due to insufficient resources, poorly trained staff, and inadequate health policy implementation. We developed a distributed GIS-based real-time syndromic surveillance system that allows collection, communication, analysis, and visualization of DE data. In the Dengue-View project, we developed a dengue surveillance, analysis, and visualization tool-set for the whole of Pakistan by employing robust and novel infrastructure to facilitate the exploration of spatio-temporal datasets that will be collected in real-time from the emergency departments (EDs) of the corresponding hospitals. The proposed system and the capabilities developed are expected to play a vital role during future dengue epidemic outbreaks (rather, any ID outbreaks) and help efficient use of the scarce resources of different governmental organizations and hospitals. Dengue-View will allow real-time monitoring of health care conditions, related to the dengue epidemic, in collaboration with the partnering hospitals. Doctors, researchers, and officials of the SERCs and the PHD, GoP, can run different filters and get a better picture of the situation and plan different preventive measures needed like insecticide spraying in certain regions or providing specific vaccination
Read MoreContact Me
Let's talk about everything!
Feel free to reach out.
Phone:
+1 541-360-9372
Email:
razaa@oregonstate.edu
08ali155@google.com
Adress:
Oregon, USA