Spring 2025 Presentation Schedule
Jan 30 - DRZ
Feb 6 - Raihan
Feb 13 - Christian
Feb 20 - DRZ
Feb 27 - DRZ
Mar 6 - Sonu (MS Project, Spring 2025)
Mar 13 - DRZ
Mar 20 - Spring Break!
Mar 27 - Fairuz (no-show!)
Apr 3 - Sonu
Apr 10 - DRZ / Reza
Apr 17 - Raihan
Apr 24 - Christian
May 1 - Fairuz
May 8 - DRZ
// DRZ on Jan. 30, 2025
Date: Jan 30 | Presenter(s): DRZ
==========================
Welcome (DRZ)
DeepSeek - AI, Stock Market, Future!?
CAPPLab website (useful info/links)
Presentation Schedule
Updates (Professional Achievements)
DRZ - four+1 grants, two+3 journals, two+4 conferences
Raihan - one+1 journals, two+1 conferences, two course projects
Sonu - MS Project report
Reza - MS Project report
Fairuz - N/A
Christian - one+1 journals, one+1 conferences
Yoel - Absent
Dr. Merve - Absent
Lab/Group Activities
Advancing Knowledge and Innovation
Collaboration and Interdisciplinary Work
Pathway to Higher Education and Careers
Technical Writing
Suggested Venues to Publish
English and Grammar
Using Template
Figures and Tables (no copy/paste!)
References (formatting)
Author Sequence
As May Arise
Send MS Thesis template to Sonu and Reza
// DRZ on Jan. 30, 2025
Date: Feb 6 | Presenter(s): Raihan and DRZ
===================================
Presentation 1: “Distributed Computing Systems for Efficient ML Model Training” by
Raihan
Summary: A distributed computing system integrates devices, edge servers, and cloud servers
to collaboratively execute computational tasks, enhancing scalability, and resource
utilization. In the context of machine learning, deploying resource-intensive models
on edge devices is challenging due to their limited computational capabilities. To
address this, efficient training frameworks are essential, involving intelligent workload
distribution, dynamic task offloading, and resource-aware scheduling to balance computations
across devices, edge servers, and cloud servers.
Questions: (i) Should we offload the full ML model to ES or partially to device and ES or partially to ES and CS? (ii) How can an ML model be effectively partitioned for improved performance and resource optimization? (iii) How can we model/simulate the problem and what optimization strategy should be used to solve the problem? (iv) Is the ML model in the devices is independent or dependent? (v) Is the training of those independent ML models in the ES are also independent or aggregated (dependent)?
Questions: (i) Should we offload the full ML model to ES or partially to device and ES or partially to ES and CS? (ii) How can an ML model be effectively partitioned for improved performance and resource optimization? (iii) How can we model/simulate the problem and what optimization strategy should be used to solve the problem? (iv) Is the ML model in the devices is independent or dependent? (v) Is the training of those independent ML models in the ES are also independent or aggregated (dependent)?
Outcomes: (i) Partially offloading to ES and CS. (ii) By applying optimization techniques.
(iii) Develop mathematical models from existing literature/optimization strategies---DNN
based and RL based. (iv) Independent. (v) Training is independent.
Presentation 2: “Simulating Wireless Network-on-Chip (WNoC) Systems” by DRZ
Summary: The presentation begins by outlining the challenges associated with WNoC systems
and provides background information on the evolution from single-core to multicore
architectures, culminating in WNoC. It then introduces a proposed approach for simulating
WNoC systems, which comprises two main steps: (i) Node Selection: identifying WNoC
nodes suitable for wireless routers, and (ii) Job Scheduling: selecting jobs and cores,
and assigning jobs to the appropriate cores. After that, the presentation delves into
simulation issues, detailing the equations used to calculate: heat indicator, hop
count, communication latency, and power consumption. Finally, it discusses questions
related to heat indicators parameters and their respective values.
Questions: (i) Are the task types acceptable? (ii) How about the task execution times? (iii) Is the equation for calculating heat indicator values suitable for this work?
Questions: (i) Are the task types acceptable? (ii) How about the task execution times? (iii) Is the equation for calculating heat indicator values suitable for this work?
Outcomes: (i) Task types are acceptable. (ii) Task execution times are adjusted per discussion.
(iii) The current equation for calculating the heat indicator is suitable for cores
executing tasks but requires modification for idle cores.
==============================
Title: “Enhancing Skin Disease Treatment: High-Speed Machine Learning Models to Boost
the Effectiveness of CADx Systems”
Summary: We have been investigating machine learning approaches to enhance the performance
of Computer-Aided Diagnosis (CADx) systems for skin disease detection. In our ACS
Omega paper, we have utilized techniques such as Generative Adversarial Networks (GAN),
Exploratory Data Analysis (EDA), and resampling methods. However, we have not captured
the models' training or inference time. While ensemble models are known to require
more time to train, they do not always guarantee improved accuracy. To develop high-speed
machine learning models, we aim to apply dimensionality reduction techniques such
as Recursive Feature Elimination with Cross-Validation (RFECV) and Principal Component
Analysis (PCA) to reduce both training and inference times without compromising model
performance.
Questions: (i) How to implement noise removal for pixel-based data? (ii) What are the main sources of error in these models when detecting skin cancer? What preprocessing methods should we apply? (iii) We have applied an ensemble CNN-SVM model and plan to apply an ensemble RF-XGBoost. Do you think there are other combinations that we can try? (iv) Could combining more than two models (e.g., RF, SVM, CNN, and XGBoost) create a better ensemble for skin cancer detection?
Questions: (i) How to implement noise removal for pixel-based data? (ii) What are the main sources of error in these models when detecting skin cancer? What preprocessing methods should we apply? (iii) We have applied an ensemble CNN-SVM model and plan to apply an ensemble RF-XGBoost. Do you think there are other combinations that we can try? (iv) Could combining more than two models (e.g., RF, SVM, CNN, and XGBoost) create a better ensemble for skin cancer detection?
Outcomes: (i) Use image-based data, and use statistical measures like mean and median to identify
and remove outliers to effectively reduce noise. (Also think about matching the pixel
data with a reference/patient-ID key and comparing it to real images to identify the
noise.) (ii) Since models can develop biases leading to incorrect predictions, explore
if autoencoders can do noise reduction and feature extraction. (iii) Implement a CNN-RF
model as it can leverage CNN's capabilities in feature extraction and RF's classification
power to improve the accuracy of CADx systems. (iv) Perform a literature review on
tri-ensemble models to understand the impact of using three or more models on classification
models (training time, inference time, accuracy, etc.).
==========================
Title: “Future Research Challenges and Funding Opportunities”
Summary: With funding cuts affecting federal agencies such as the NSF and NIH, it is essential
to explore grant opportunities from industry that bridge academic research with real-world
technology challenges. The Cisco Distributed Systems, SONY Faculty Innovation Award,
Alternatives Research & Development Foundation (ARDF)'s Annual Open Grant Program,
and Samsung Research America (SRA) Strategic Alliance for Research and Technology
(START) programs are some promising avenues for funding. Key AI/ML focus areas include
deep learning, on-device AI, generative AI, and multi-modal machine learning. Our
goal is to submit proposals that deliver high-performance, real-time, power-efficient,
and always-available solutions. To achieve this, researchers are developing AI/ML
models that require minimal disk space, runtime memory, and computational resources.
They are also advancing the field by designing deep learning models compact enough
to run on-chip with ultra-low power consumption, memory usage, and latency.
Questions: (i) What/How can we apply to Cisco Distributed Systems Program? (ii) What/How can we apply to SONY Faculty Innovation Award Program? (iii) What/How can we apply to ARDF's Annual Open Grant Program? (iv) What/How can we apply to SRA START AI Program?
Questions: (i) What/How can we apply to Cisco Distributed Systems Program? (ii) What/How can we apply to SONY Faculty Innovation Award Program? (iii) What/How can we apply to ARDF's Annual Open Grant Program? (iv) What/How can we apply to SRA START AI Program?
Outcomes: (i) Raihan will work with DRZ and collect results for the Cisco (Distributed Systems)
proposal. (ii) Christian will work with DRZ and collect results for the SONY (Biomedical
and Life Science | Cell Biology | AI-assisted High-speed Cell Image Analysis) proposal.
(iii) DRZ will submit the ARDF LOI after discussing it with Christian. (iv) Fairuz
will work with DRZ to identify challenges/topics for Samsung's (Artificial Intelligence
| Multi-modal life-long memory augmentation system with fast retrieval) proposals.
==========================
Title: “Future Research Challenges and Funding Opportunities”
Note: Continue our discussion on this tipoc.
============================
Title: “Predicting Performance of Heterogeneous Edge-Cloud Systems Using Machine Learning
Models”
Summary: Edge-cloud systems consist of heterogeneous computational infrastructures designed
to efficiently manage distributed workloads. Accurate performance prediction is essential
for optimizing resource allocation, reducing latency, and improving system efficiency.
This study applies machine learning models—Random Forest (RF), Long Short-Term Memory
(LSTM), Deep Neural Networks (DNN), Recurrent Neural Networks (RNN), and a hybrid
RNN-DNN model—to predict key performance metrics. Using performance datasets, these
models are evaluated based on Mean Absolute Error (MAE) and Root Mean Square Error
(RMSE). Results show that RF achieves the highest predictive accuracy, while deep
learning models like LSTM and DNN yield varying effectiveness. The hybrid RNN-DNN
model balances complexity and accuracy. These findings underscore the potential of
ML-based approaches for optimizing resource management and performance in cloud-edge
architectures.
Questions: (i) What additional performance metrics could provide deeper insights into model effectiveness? (ii) How can hyperparameter tuning further enhance the accuracy of deep learning models in this context? (iii) Would alternative ML approaches, such as ensemble learning or transformer-based models, improve predictions? (iv)What are the real-world deployment challenges of these models in edge-cloud environments?
Questions: (i) What additional performance metrics could provide deeper insights into model effectiveness? (ii) How can hyperparameter tuning further enhance the accuracy of deep learning models in this context? (iii) Would alternative ML approaches, such as ensemble learning or transformer-based models, improve predictions? (iv)What are the real-world deployment challenges of these models in edge-cloud environments?
Outcomes: (i) TBD.
============================
No Presentation - DRZ was sick.
// DRZ on March 15, 2025
Date: March 20 | Spring Break!
Date: March 27 | Presenter(s): Fairuz
==============================
No Presentation - Fairuz did not present and did not provide prior notice.
// DRZ on March 27, 2025
Date: April 3 | Presenter(s): Sonu
===========================
Title: “Predicting Performance of Heterogeneous Edge-Cloud Systems Using Machine Learning
Models”
Second Presentation.
// DRZ on April 10, 2025
Date: April 10 | Presenter(s): DRZ / Reza
=================================
Presentation 1: Judge Rachel Pickering's slides on "Artificial Intelligence: A Look
Back to See Its Future" presented at the 2025 Wichita Council of Engineering Societies
(WCES). DRZ collected the slides from Mark Hunter of IEEE Wichita Section to share
with students.
Summary: The presentation starts with defining AI and related history, including ChatGPT and
Deepfake. Important topics include: how we are using AI in the law, thoughts of the
future with AI, Quantum computing, etc. The presentation ends with: "Without accuracy,
accountability, confidentiality, and privacy, AI advancements fall flat."
Questions: (i) Are you aware of the AI related legal issues/laws? (ii) What is your thought on the future of AI? (iii) Accuracy is important. How about accountability, confidentiality, and privacy for AI advancements?
Questions: (i) Are you aware of the AI related legal issues/laws? (ii) What is your thought on the future of AI? (iii) Accuracy is important. How about accountability, confidentiality, and privacy for AI advancements?
Presentation 2: “Handwritten PDF to Excel Conversion: Evaluating Automation Tools
for Accuracy and Efficiency” by Reza
Summary: Automating the conversion of handwritten documents—typically received as PDF or JPEG
files—into Excel format significantly enhances efficiency and accuracy in data processing.
This reduces manual effort, minimizes errors, and supports faster decision-making.
This study evaluates artificial intelligence (AI)-powered tools, including CamScanner,
Online OCR, Docsumo, Amazon Textract, and Microsoft Power Automate. Power Automate
stands out with 98.9% accuracy due to its fixed-format model training, email-triggered
automation, and seamless Microsoft integration. In comparison, CamScanner and Online
OCR achieved 51.04% and 7.29%, respectively. These findings highlight the value of
AI tools in streamlining workflows and improving data integrity across sectors like
healthcare, banking, and education.
Questions: None.
Questions: None.
Outcomes: TBD.
=============================
Title: “Deep Learning-Driven Task Scheduling Optimization for Enhanced Performance in Edge-Cloud
Heterogeneous Systems”
Summary: This work proposes a deep learning-driven task scheduling framework for heterogeneous
systems based on the edge cloud to improve performance. A heterogeneous system comprising
40 Internet of Things (IoT) devices, one cloud server (CS), and four edge servers
(ESs) is evaluated with varying task sizes from 2 GiB to 10 GiB. Each ES is employed
with a Deep Neural Network (DNN) to optimize task scheduling and simultaneously minimize
execution time, energy consumption, and CS utilization. Comparative analysis demonstrates
that our approach outperforms the Optimal Pairing Ratio (OPR) and traditional cloud-only
methods.
Questions: (i) What system states (observations) should we include in our RL environment to best capture edge-cloud dynamics? (ii) What kind of reward function would most accurately reflect training quality versus resource usage trade-offs? (iii) How can we simulate realistic edge-cloud latency and GPU performance to evaluate the RL agent effectively? (iv) Any suggestions/recommendations?
Questions: (i) What system states (observations) should we include in our RL environment to best capture edge-cloud dynamics? (ii) What kind of reward function would most accurately reflect training quality versus resource usage trade-offs? (iii) How can we simulate realistic edge-cloud latency and GPU performance to evaluate the RL agent effectively? (iv) Any suggestions/recommendations?
Outcomes: (i) Apart from those already listed in the slides, we can also include one more "Edge
GPU available cores." (ii) We start coding with the existing reward function, and
then based on the output, we will modify the reward function if necessary. (iii) Using
the papers recommended by Fairuz to model mathematically and simulate using Python
in Google Colab. (iv) In the reward function, consider server utilization instead
of edge GPU utilization. Be specific to the LLM Model (i.e., GPT 2/GPT 3) while splitting
the model and collecting the results.
===============================
Title: “Evaluating the Impact of Computing Platforms and Optimization Techniques on
Machine Learning Model Training and Inference Time"
Summary: This study evaluates how different computing platforms—an HPC cluster, desktop, and
laptop—affect the performance of seven ML models, including CNN, RF, XGBoost, and
ensemble approaches. After standard preprocessing with SMOTE and RFECV, we measured
training time, accuracy, and inference time across platforms. Surprisingly, the desktop
consistently outperformed the HPC and laptop in both efficiency and predictive accuracy,
challenging common assumptions about HPC superiority and highlighting the importance
of system-level optimization.
Questions: (i)What novelty or unique contribution can we emphasize to strengthen this paper's impact? (ii) What other metrics or system factors should we consider? (iii) Are there further optimization strategies we should explore?
Questions: (i)What novelty or unique contribution can we emphasize to strengthen this paper's impact? (ii) What other metrics or system factors should we consider? (iii) Are there further optimization strategies we should explore?
Outcomes: (i) TBD.
===========================
Title: “Inference Time Optimization Techniques in ML Models"
Summary: Inference time in machine learning is becoming increasingly important with the growing
popularity of AI applications. Inference time refers to the time a model takes to
generate an output for a given input. As the need for faster and more accessible inference
grows, running ML models directly in the browser is gaining attention. Technologies
like WebAssembly and WebGPU play a key role in this shift by enabling parallelism
and batching, significantly improving inference performance. Recent developments on
inference time will be discussed.
Questions: (i) Does this area have enough potential to be explored, and do we have the resources to do so? (ii) What could be a potential novel idea to explore in this topic? (iii) Any ideas / suggestions / recommendations?
Questions: (i) Does this area have enough potential to be explored, and do we have the resources to do so? (ii) What could be a potential novel idea to explore in this topic? (iii) Any ideas / suggestions / recommendations?
Outcomes: (i) Further study is needed on inference time and related developments.
==========================
Title: “End of Semester Updates"
Summary: CAPPLab researchers discuss their activities and achievements in spring 2025 semester
and plans for summer and fall 2025 semesters. DRZ will share Judge Rachel Pickering's
slides on "Artificial Intelligence: A Look Back to See Its Future" presented at the
2025 Wichita Council of Engineering Societies (WCES).
Questions: (i)What are your plans for summer and fall 2025 semesters?
Questions: (i)What are your plans for summer and fall 2025 semesters?
Outcomes: (i) TBD.