A MATLAB-based ray tracing simulator for modeling millimeter-wave signal propagation in indoor environments. The framework traces rays from a transmitter through a configurable 3D office layout, computing received power while accounting for free-space path loss, multiple reflections (up to 3 bounces), and log-normal shadowing. Features include color-coded power visualization and animated GIF generation.

A deep-learning-based traffic pattern detection system using multi-modal data from radio signals and camera images. The system captures real-time RF signals via a USRP and synchronizes them with camera images to classify traffic patterns into Very Busy, Moderate, and Not Busy states using a Complex-Valued Neural Network (CVNN), achieving over 80% accuracy with sub-100ms latency.

We test the EdgeSLAM algorithm for generating a map to ensure that it can be used at a construction site building where the aim of the robot is to navigate to specific goals while avoiding humans and maintaining a safe distance of 1.25 meters from the human workers at all points in time. Further, our approach uses a YOLOv3 network to detect a human worker and then performs clustering and Kalman filtering for point-cloud data from a depth camera to estimate the relative human motion under the robot’s local coordination

This project focuses on text classification using a BERT-based model on a Spanish news dataset. The dataset is preprocessed and split into training, validation, and test sets. The model architecture includes input layers for tokenized text and attention masks, followed by a BERT layer and a softmax output layer. The model is trained using TensorFlow with Adam optimizer and categorical crossentropy loss. Visualization techniques, such as training and validation accuracy plots, training and validation loss plots, and a confusion matrix, are employed to analyze the model's performance.

Markov model of the Voynich Manuscript provides us with useful information for analysing the Voynich script. We used the transition matrices generated by our model to compare the six sections of the Manuscript, to compare the Manuscript to several natural languages, and to calculate the most likely character for several unknown characters in this transcription of the Voynich Manuscript.

This project centers on the re- implementation of the Vision Transformer (ViT) architecture, a pivotal development in computer vision. ViT utilizes self-attention mechanisms for image recognition, promising to surpass conventional convolutional neural networks (CNNs). Challenges to be addressed include scaling ViT for larger images and datasets while maintaining efficiency, devising efficient pretraining strategies in data-limited scenarios, optimizing fine-tuning techniques for specific tasks, and enhancing interpretability.

The Hand Gesture Recognition project utilizes the Sign Language MNIST dataset, a modification of the classic MNIST, containing 27,455 training cases and 7,172 test cases representing 24 classes of American Sign Language letters. The dataset was expanded using image processing techniques to create variations for better machine learning model training. This project aims to develop a robust visual recognition algorithm to assist the deaf and hard-of-hearing in better communicating through computer vision applications

This project, conducted in collaboration with a partner, focuses on implementing architectures for image classification using the CIFAR-100 dataset. The architectures consist of convolutional layers with pooling and at least one fully connected layer, followed by softmax for the output layer. TensorFlow is utilized for model implementation, with a main loop for training defined in separate Python files. Results and analysis are presented in a written report following a provided template, including experimental setup, results, and conclusions.

This project explores the design and implementation of two architectures for image classification using the Fashion MNIST dataset. The architectures consist of fully connected layers with ReLU activation for hidden nodes and softmax for the output layer. Hyperparameter selection is conducted independently of the test set, with validation data obtained through either a single training set and a single validation set or k-fold cross-validation. The models are optimized using Adam, with at least two sets of hyperparameters and one regularizer evaluated for each architecture.

This project compares recurrent and attention-based architectures for sequence modeling tasks. We implemented and evaluated GRU (Gated Recurrent Unit), LSTM (Long Short-Term Memory), and Transformer models on time-series prediction and natural language processing tasks. The study analyzes the trade-offs between computational efficiency, long-range dependency handling, and model performance across different sequence lengths and complexity levels.