Alireza Torabian

• In the field of Conformal Prediction.

• Exploring the usefulness of various conformity measures in obtaining more informative prediction sets and confidence intervals.

• Developed a plugin that will obscure images in order to increase privacy that achieved 35% success rate.

• Performed a black‐box adversarial attack on facial recognition using projected gradient descent with momentum in latent space of FaceNet (an Inception-Resnet model).

• In face detection attack, the overlap between the detected areas by SSD MobileNet V1 and the actual faces is minimized using PGD.

• Image augmentations are used to apply the attacks on black-box models.

• Developed a Persian question answering system in Python Tensorflow.

• LDA is used to cluster similar questions. To expand our dataset, we utilize the answers to questions in the same cluster interchangeably.

• LSTM Seq2Seq model with Luong-style attention mechanism is used to generate an answer to a question.

• Object detection task is performed to detect victims using the YOLO (You Only Look Once) algorithm. YOLO has a lower coverage of bounding boxes compared to SSD (Single Shot Multibox Detector), but it is faster, which is important for our real-time application.

• Developed an autonomous exploration algorithm and path planner for robots to help them explore a map simultaneously.

• Co-founder and a member of Dropout (F.K.A. Lost Soul) games. Developing android games using Java in Android Studio.

GPA: A+
Advised by Dr. Ruth Urner

GPA: 3.9/4 (18.25/20)
In the top 10% among 101 students

Thesis: Design and Implementation of an Automatic Question Answering System for Answering Simple Persian Questions
Advised by Dr. Saeedeh Momtazi
THESIS

• A graph autoencoder has been applied to an actor's network to map the actors to a latent space, using Keras in Python.

• Achieved 99.46% accuracy on link weight prediction task for weights between 0 and 1.

• An alternative actor is found by searching the latent space using a K-d tree, and a co-star is suggested based on the predicted weights in the autoencoder model's target network.

• ML Reproducibility challenge of the paper "Autoencoding Keyword Correlation Graph for Document Clustering".

• Applying a graph autoencoder to a graph of keywords to map each document to a latent space which is used for clustering.

• My bachelor thesis that was building a question answering system based on a knowledge-base in Python.

• SVM and CNN classification models used to classify the question type achieved 96% accuracy and F1‐score of 92.7%.

• Implementing algorithms to optimize a convex problem.

• Some unconstrained optimizations such as line search and trust region methods, and some constrained optimizations such as log barrier.