THE PORTFOLIO OF BERNARDO BIANCO PRADO

Email
Linkedin
GitHub

HUMAN POSE ESTIMATION WITH VISION TRANSFORMERS
In this project, we implemented vision transformers (ViTPose) and convolutional pose machines (CPM) for the task of pose estimation. We trained our model on the COCO dataset and experimented with the number of layers in the vision transformer model to see if we could get good results with a smaller scale-model. For CPM, we achieved a mAP of 59.5, with a half-sized ViTPose model, we achieved a mAP of 67.1, and with a full-sized ViTPose, we achieved a mAP of 71.0. Curious visitors can check out our project report.




SHOE RECONSTRUCTION USING GENERATIVE ADVERSARIAL NETWORKS
Implemented and trained, in PyTorch, a PatchGAN network to reconstruct shoes from drawings. Below are two examples from the training data and one example of the reconstruction of a shoe I drew.




STYLE TRANSFER USING SQUEEZE NET
Applied a PyTorch SqueezeNet network to turn pictures into paintings. Below are two examples.




SCALABLE AND EFFICIENT MACHINE LEARNING ALGORITHMS
This repository features code that was written for a graduate course at the University of Michigan which focused on efficient and scalable optimization algorithms for machine learning. The problems solved include logistic regression, robust principal component analysis, matrix completion, dictionary learning and orthogonal dictionary learning.


PREDICTION OF BREAST CANCER SURVIVAL USING MACHINE LEARNING
This project uses machine learning algorithms to predict the survival of patients 10 years after they are diagnosed with breast cancer. In addition, we sought to understand what are the most important parameters in the dataset for prediction of survival. The machine learning algorithms were implemented by ourselves and the methods used were random forests, L1 logistic regression and kernelized support vector machines. Our most effective methods achieved around 80% prediction accuracy.

PARALLELIZED RAY TRACER This project implements Peter Shirley's Ray Tracing in One Weekend in C++, but it differs from the book's implementation in that the tracer was parallelized using CUDA. The parallelized code achieved a performance improvement of 12 times compared to the original code!




HUNGRY CAT COOKING Hungry Cat Cooking is a cooking platforming game where you will play as a cat that aspires to be the first feline to open a human restaurant! Of course, in the world we live in, a cat can't just randomly open a restaurant and serve humans. A long time ago, a fellow feline tried to accomplish such feat and wasn't successful. To your luck, this old cat created a training that will prepare you to thrive where he has failed. Check out the game trailer below!



Developed by: Bernardo Bianco Prado, Calvin Tran, Colin Yoon, and Drew Vanderspool
Original scores by: Emily Sulkey and Liam Kochanek
Original art by: Liam Kochanek, Calvin Tran and Bernardo Bianco Prado
Voice acting by: Chandler Branson


PIXEL BLUE This is a two week prototype of a rythmic puzzle game in homage of Blues music. Your one and only tool is an electric keyboard and you can use it to progress in the game, interact with the world around you or just enjoy your own Blues improvisation!


THE LEGEND OF ZELDA (UNITY REMASTER)

H. V. Jain, K. A. Norton, B. B. Prado, T. L. Jackson. SMoRe ParS: A Novel Methodology for Bridging Modeling Modalities and Experimental Data Applied to 3D Vascular Tumor Growth. Frontiers in Molecular Biosciences Biological Modeling and Simulation, 2022. [journal]






B. B. Prado, K. Klinger-Logan. Linear Operators, the Hurwitz Zeta Function and Dirichlet L-Functions. Journal of Number Theory, 2020. [journal] [arxiv]

The Global Landscape of Neural Networks
This article reviews some mathematical and computational results that seek to understand when neural networks converge to an optimal or nearly optimal output.


A Survey of Methods for Streaming Principal Component Analysis
This article surveys different methods that address the computational and data-related challenges of principal component analysis (PCA). The issues highlighted are that PCA can be computationally expensive and that in practice, most of the data gathered is complete (e.g. has many missing entries). We review a series of methods that seek to address these issues.