The Speakers

Rania is an Assistant Professor in the Computer Science department and the co-organizer of this conference. The title of her talk is Enhancing Efficiency in Embedded Machine Learning: Techniques for Neural Network Deployment.

Abstract: Deploying neural networks on embedded systems poses significant challenges due to their limited computational capabilities and strict energy constraints. In this talk, I will present three key techniques that address these challenges: approximate multipliers, quantization, and pruning. Approximate multipliers offer a smart balance between efficiency and accuracy, allowing for reduced energy consumption and hardware simplicity. Quantization, by lowering the precision of model parameters, cuts down on memory usage and speeds up inference, making it ideal for real-time applications. Pruning goes a step further by eliminating unnecessary parameters. Through a blend of these strategies, we can push the boundaries of what is possible in embedded machine learning, enabling more powerful and efficient applications. 

Steven is a Mathematician Scientist, educator and editor at ZbMath_OPEN. His talk is titled  Learning Manifolds in high dimensional Euclidean Space and Whitney Extensions.

Abstract: One of the main challenges in high dimensional data analysis, AI, DL and many other areas is dealing with exponential growth of the computational and sample complexity of several needed generic inference tasks as a function of dimension, a phenomenon termed "the curse of dimensionality".
One intuition that has been put forward to lessen or even obviate the impact of this curse is a manifold hypothesis that the data tends to lie on or near a low dimensional submanifold of the ambient space. Algorithms and analyses that are based on this hypothesis constitute an enormous area of research of deep learning theory known as manifold learning. One may, under certain frameworks, view the manifold hypothesis as a Whitney extension problem whose statement is the following:

Given a finite set $E$ in $\mathbb R^d, d>1$ and a function $f:E\to \mathbb R$. How to understand if there exists a function extension $F:\mathbb R^d\to \mathbb R$ which agrees with $f$ on E$ and is smooth.

The talk will focus on learning manifolds and their connections to Whitney extensions. This body of work is in a recent book by the author:

Reference:

1. S. B. Damelin, Near extensions and Alignment of Data in R^n: Whitney extensions of smooth near isometries, shortest paths, equidistribution, clustering and non-rigid alignment of data in Euclidean space, John Wiley & Sons 2024.

Tanweer is a Professor in the Math & Stats department at EMU. Her talk is titled  On some machine learning techniques with applications to statistics.

Abstract: Machine learning (ML) techniques utilize data experiences by learning through model building processes and make improved decision thereby, without being explicitly programmed. This presentation provides some applications of ML techniques such as Support Vector Machine, k-Nearest Neighbors (KNN), Logistic Regression Model, Discriminant Analysis, etc., which are of great use in statistical decision-making processes. We seek to apply ML techniques for analyzing real-life data by using R programming language.

Samuel is a student at Adrian College. He will talk about Comparative Analysis of Machine Learning Algorithms for Predictive Statistics on Lung Cancer Patients.

  Abstract: Lung cancer is an ever-growing issue throughout the world that requires sophisticated tools for effective analysis. In this comparative study, three machine learning algorithms will be used to analyze a dataset concerning lung cancer patients to perform predictive statistics upon said patients to test the effectiveness of the used algorithms. To explain, machine learning is a subset of artificial intelligence that refers to machines or systems that are capable of adapting based on prior decisions made to increase accuracy and to yield a consistent and accurate output. Using Python as well a certain machine learning module that can be used with Python, the data will be processed, analyzed, and reported on. With that said, the machine learning algorithms will be compared using certain performance metrics, those being the accuracy, precision, recall, F1 score, and the Area Under Curve of the Receiver Operating Characteristic (AUC-ROC) of each algorithm. Furthermore, each of these algorithms will differ and certain ones will be better fits for this project than others. This research may prove itself to be valuable as cancer is a constant issue in the present day and it will show the effectiveness in using machine learning algorithms to analyze data in the modern age. Beyond this, the research holds implications to the future as machine learning continues to be developed, which may eventually allow researchers to find a method to significantly reduce deaths related to cancer in the future. 

Siyuan is an Associate Professor in the Computer Science department at EMU. Siyuan will talk about  Contrastive Learning for Cross-Language Function Embeddings in Code Search .

  Abstract: Pretrained language models for code token embeddings are used in code search, code clone detection, and other code-related tasks. Similarly, code function embeddings are useful in such tasks. However, there is no out-of-box models for function embeddings in the current literature. So, this paper proposes CodeCSE, a contrastive learning model that learns embeddings for functions and their descriptions in one space. CodeCSE will be open-sourced, and is uploaded to the HuggingFace public hub. We evaluated CodeCSE on code search. CodeCSE’s cross-language zero-shot approach is as efficient as the models finetuned from GraphCodeBERT for specific languages. 

Fei is a Research Fellow in EECS at the University of Michigan. He will talk about Enhancing Robustness in Deep Learning: Insights from Unsupervised Domain Adaptation, Compound Domain Challenges, and Large Language Models.

  Abstract: In recent years, deep learning models have demonstrated remarkable performance across various AI applications, particularly in domains such as semantic segmentation and attribute extraction. However, the robustness of these models to domain shifts and data scarcity remains a significant challenge. This presentation addresses the issue of model robustness through an exploration of three aspects. 1) I will introduce the self-supervised learning can be used for unsupervised domain adaptation, with an example from semantic segmentation task. I will present our work on unsupervised intra-domain adaptation for semantic segmentation, aiming to bridge the gap between synthetic and real-world data. 2) I will present the domain adaptation challenges in real applications such as autonomous driving, with a focus on compound domains. I will present this part with our paper on open compound domain adaptation for semantic segmentation. 3) I will present the generalization of ML models, specifically in the era of Large Language Models (LLM). I will introduce our newest research on zero-shot workflow by utilizing large-scale vision and language models to reduce reliance on human annotations. Through these discussions, this presentation highlights the importance of robust deep learning models in overcoming challenges posed by domain shifts, data scarcity, and generalization. Insights from these papers offer valuable contributions towards advancing the robustness and adaptability of deep learning models in real-world applications. 

Jason is a student at Adrian College. He will talk about Sign Language Hand Gesture Recognition using Google's Media Pipe. Jason works under the supervision of Dr. Yasser Alginahi.

  Abstract: Sign Language Recognition (SLR) is a technology that enables computers to identify sign language gestures. There are different methods for achieving SLR, each with its own advantages and disadvantages. This discussion focuses on the use of real-time image processing and machine learning algorithms to detect American Sign Language (ASL) gestures. For the acquisition and detection of hand gestures, Google MediaPipe was utilized in this project. The dataset was generated using Google's Teachable Machine. Teachable Machine offers a user-friendly interface that simplifies the collection of image data through a video camera. The data for each class is organized into separate folders, making it easy to import and train the model. Google's MediaPipe framework is a recent development that enables efficient hand gesture recognition, even with small datasets, by abstracting computer perception data. MediaPipe breaks down computer perception into components that can be combined to provide inference from a trained machine learning model. This project utilizes MediaPipe to detect hand gestures from the ASL alphabet. The model incorporates approximately 400 samples per class, with 26 classes representing each letter. In a real-time application, the accuracy achieved by the model is 83%. The application provides instant feedback on the corresponding ASL hand gesture that has been detected. To accomplish hand detection and gesture recognition, a customized "Hand Gesture Recognition Task" model is employed. 

Katherine is a student at Adrian College. She will talk about A Survey of Deep Learning Computer Vision Algorithms in Mobile Robotics. This work is done under the supervision of Dr. Yasser Alginahi.

  Abstract: Computer vision is one of the many areas of machine learning that is seeing advancements due to the increased prevalence of deep learning, which in turn presents solutions to many questions in robotics. Traditionally, mathematical algorithms and transforms have been used to emulate object recognition in computer vision systems, but deep learning has introduced biologically-inspired artificial neural networks for solutions in this area. This rapidly evolving technology has played an integral role in the improvement of both robotics vision and long-term autonomy, promising myriad benefits for humanity. In this project, various deep learning models will be applied and deployed for computer vision tasks on a mobile robot, with the goal of finding the most favorable and efficient results. A Raspberry Pi 4 will be used with a robot kit to create a mobile robot that has two main functions: to be able to safely traverse its environment while avoiding obstacles, and to be able to find and identify certain objects. 

Brian is a computer science student at EMU. He will talk about AI Text Classification Using Ensembled Transformer Models .

  Abstract: With the widespread proliferation and ease of access of large language models like ChatGPT, the need to determine whether text is AI generated or human written is now a more critical task than ever. This research outlines and compares several different state of the art methods used and attempts to synthesize them in order to propose an algorithm that improves upon existing methods, paving the way for improvements over currently available commercial methods that have proven inadequate in real world contexts. 

Andrew is a Professor in the Department of Math & Stats. He will talk about Adapting Existing Machine Learning Models for Efficient Biochemistry Imaging Analysis.
This is joint work with  his student Emily Marron.

 
  Abstract: A Case Study at Eastern Michigan University Biochemistry researchers at Eastern Michigan University need to segment transmission electron microscope images to determine the size of autophagic bodies in yeast cells. This task is quite time-consuming to do by hand, so they wanted help from a machine learning (ML) model. Building a new ML model specifically for this would be difficult, especially given our small dataset of just over 250 images. So, we tried to adapt an existing model to our needs. We found Cellpose 2.0’s CPx model to be the closest match. After training this model with our data, it performed as well as our researchers did. This model can be used by itself for image analysis, or its output can be double checked by a human expert, which is still 5 times as fast as having the human expert do it alone. This approach illustrates that it is not always necessary to "reinvent the wheel"' when it comes to machine learning, and that many existing solutions can be adapted with relative ease in a short time frame. The entire project (plan, build, and deploy) was done over the Summer of 2023. For more details, see https://www.biorxiv.org/content/10.1101/2023.10.23.563617v1

Mohammad is a Math and CS student at EMU. His talk is about  CodeContext: Integrating External Context for Enhanced Source-Code Model Performance in Software Development.

 
  Abstract: In the ever evolving field of software development, understanding and maintaining complex codebases is crucial. Existing source-code machine learning models aid this, yet they often overlook an important factor: the code's context. Our research focuses on leveraging external contextual information to enhance source-code model performance. We’ve developed a data pipeline that utilizes CodeQL to extract contextual information from the CodeSearchNet dataset and developed CodeContext, a transformer model that integrates context with code in an effective manner. This approach promises to enhance code comprehension and maintenance, marking a significant advancement in software development tools. 

Matthew is a student at Adrian College working under the guidance of  Dr. Yasser Alginahi. His talk is about  Harnessing Machine Learning and Neural Networks for English Alphabet Letter Recognition.

 
  Abstract: The advancement of artificial intelligence technology is rapidly transforming various fields of study. To understand how computers interact with their surroundings, it is crucial for people to grasp this concept. This presentation will demonstrate a deep learning program that utilizes neural networks to read alphabetical letters. The program will be implemented using Python programming and PyTorch, a machine learning framework for computer vision and natural language processing. It will then be executed on the Raspberry Pi computer board. The main objective of this presentation is to create software that can accurately recognize and process the different letters of the English alphabet. During the testing phase, the program's recognition capabilities will be assessed, with a strong emphasis on achieving high accuracy. By attending this presentation, you will gain valuable insights into the training processes, extensive dataset handling, and informed decision-making capabilities that deep learning and neural networks enable computers to possess. 

JaSai is a Computer Science student at EMU. His talk is about  Human Actions Through Computer Vision Insights .

 
  Abstract: In the realm of computer vision, deciphering human actions stands as a formidable challenge. This talk will explore the field of human action recognition through the lens of computers. By leveraging machine learning algorithms and deep neural networks, this research explores the intricate process of interpreting human actions from visual data. Utilizing advanced video and image processing techniques, we can extract valuable insights to classify various human movements accurately.