Shafin Haque

Hi! I'm Shafin, a freshman at UC Berkeley studying Electrical Engineering and Computer Sciences (EECS).

I'm interested in machine learning and computer vision. This past summer I interned for Invisible AI's machine learning team, where I built deep learning computer vision systems for improving automotive manufacturing lines.

Previously, I was researching for the the Vision Research Lab at UC Santa Barbara, working on objection detection methods for ROV marine data.

Check out my work and experience on this page, and feel free to reach out!

Email  /  Resume  /  LinkedIn  /  Google Scholar  /  Github  /  Devpost

My research primarly focuses on computer vision, but I have also worked in other deep learning fields.

We propose Box Prediction Rebalancing for single-stage object detectors to combat the challenge of learning from partially annotated datasets. By randomly removing percentages of negative predictions from our model's loss computation, our model performs better as it learns less from false positives which may be true species of interest without ground truth.

We introduce a novel dataset for detecting car accidents and unusual street activity through live CCTV cameras. We annotate the videos with metadata to help with future trend prediction, as well as give more information for each video. We also propose a deep learning system to cluster similar activity based on latent space representations.

We propose a novel GAN structure for semi-supervised classification of medical images in restricted, fully supervised data. We incorporate a classifier into the adversarial relationship such that the generator trains adversarially against both the classifier and discriminator.

In an effort to streamline the process for medical practitioners and improve clinical care for patients within the hospital we introduce a machine learning based NLP approach for classifying healthcare conditions from transcription notes.

I have been fortunate enough to work for both companies and universities solving machine learning problems.

• Developed custom few-shot deep learning model for semantic segmentation of workpiece objects in automotive manufacturing lines. Trained segmentation model with only 15 images per camera site.
• Created light-weight single-shot detector head using residual skip connections from existing production pose estimation model as the backbone to detect worker faces for face blurring in live cameras.

• Worked with MARE Group to develop dataset for streamlining identification and counting process of marine species in underwater ROV videos. Researched novel object detection methods to improve performances on our dataset.
• Wrote extended abstract accepted to NeurIPS LMLR proposing novel bounding box prediction rebalancing method for single-stage object detectors such as YOLO to improve learning from partially labeled datasets.
• Also contributed to research paper accepted to International Journal of Computer Vision and CVPR Workshop.

• Researched the application of federated learning in medical imaging in order to improve security and accuracy of 2D and 3D segmentation of multi-organ and tumor abdominal CT scans.

• Programmed deep learning models in TensorFlow and PyTorch to detect distracted/impaired drivers in images and video.
• Used custom and pre-trained CNNs to reach an accuracy of 99% for distinguishing between multiple classes of distracted and non distracted driving. Also tested with CNN-RNN architecture for video classification.
• Lead computer vision team on the project by overseeing all team operations and holding meetings. Lead collaboration with software engineering team to develop mobile application using Flutter.

I've competed in numerous hackathons and these are some of the projects created during them. There are also some personal projects listed and more of my projects can be found on my Github or Devpost.

• Python web app for prediction various characteristics of wildfires with ML. Used neural networks, regression, and bounding box detection models
• 4th place Congressional App Challenge 2021 & 4x Hackaton Award Winner

• RShiny app for predicting the next pitch in baseball based on current game scenario with R and Python
• Wrote article published in TowardsAI documenting process

• Flask app for detecting eye disorders and effects of treatment with Firebase backend
• 4x hackathon award winner

• Vue.JS app with TensoFlow.js and AutoML for detecting and classifying COVID-19 and the flu with Firebase backend
• Google sponsor prize winner

• Web app for detecting and predicting various characteristics of different natural disasters
• 6x hackathon award winner

Website template from here