Marco Alberto Hojel

I am a machine learning researcher and senior Honors CS student at UC Berkeley.

I learned a lot about panoptic segmentation of satellite images, informal urban sprawl, and entrepreneurship after spending almost a year obsessing over an idea, advised by Grant Ricketts and Alejandro Maldonado, incubated at SkyDeck, receiving the Berkeley's Changemaker Award at the Big Ideas Competiton and support from AIEB, only to realize I lacked the political maneuverability and intuition to deploy it in my home country, Mexico.

I am now going back to first principles to develop my skills and intuitions as a creator, developer, and leader.

Email  /  Twitter  /  LinkedIn  /  GitHub

Aligned with David Deutsch's views, I believe explanations to be the purest form of human achievement.

We explore the activations of different attention heads in a visual prompting model, MAEVQGAN, and consequently leverage learned representations of tasks to guide the behaviour of the model ultimately improving performance and speed. To do this, we utilize the REIFNORCE algorithm to find optimal patching positions where we clamp the output of particular heads to their mean activations when prompted with the intended task.

Advised by Gireeja Ranade, we explore previous work in developing provably robust deep neural networks that can defend themselves against adversarial attacks. In the main paper, we explore the use of a convex outer polytope of an input's L-2 norm ball after going through the network to identify whether there are any changes in classification within the outer bound. In the supplemental material, we explore the use of FGSM vs box-constrained L-BFGS in finding adversarial examples.

Advised by Umesh V. Vazirani, we document our approach to building and benchmarking a Variational Quantum Classifier (VQC) in Qiskit to solve a simple two-dimensional binary classification task with synthetic data. By encoding the training data into the amplitudes of quantum states through a Zero-Pi-Pulse (ZZ) feature map--essentially a parametrized circuit that applies controlled-Z gates to an input quantum state based on the input classical data--and training an ansatz circuit implemented as a Real Amplitudes circuit which manipulates the input quantum state through unitary evolution to minimize the loss.

Advised by Michael Larkin, I delve into the intricacies of the failing property registry in Mexico and make an argument for how satellite imagery could serve as a tool to turn this around. From a macroeconomic analysis to a survey of the relevant political environment, this research paper identifies and segments the deeply-rooted vicious cycle that Mexican society has suffered from since the Hispanic era. By leveraging deep learning methods for geospatial data analytics from remotely-sensed images, local Mexican governments could observe and detect informal urban sprawl and take steps to set up appropriate infrastructure from the get-go.

Advised by Dr. Daniel Kammen, Professor of Energy at Berkeley and contributing author to Nobel Peace Prize-winning IPCC, and Chris Jones, Director of the CoolClimate Network, I conducted analysis on consumption-based greenhouse gas accounting, contributing to a modular and serverless web application built on Firebase and ReactJS. Synthesized existing U.S. datasets on climate change, greenhouse gas (GHG) emissions, GHG drivers, household carbon footprints, climate action planning, climate impacts, adaptation, pollution, environmental justice, and individual behavior into a single, data-driven climate action portal called EcoDataLab for U.S. cities and communities.

I feel most fulfilled when developing impactful solutions to real-world problems or when messing around with inspiring new tech.

After presenting a demo of a semantic-search-powered course content repository I built for a class--EECS16B, where we teach students everything that comes behind building a voice-controlled electric vehicle from band-pass filters to control theory to intermediate linear algebra--I was TA-ing at Berkeley for Cerebral Valley's AI Coworking session attendees at Shack15, Bryan and Kathy approached me to talk about their vision for in-meeting AI assistants. We quickly hit it off and I ended up building out their cloud infrastructure to stream Zoom transcripts into GPT-4 to extract key insights and action items.

This was my first serious attempt at entrepreneurship. During my time as the founder of BlackPrint Technologies, I sought to confront the multifaceted and technically challenging problem of updating Mexico's property registry. By leveraging AI to extract geospatial data, specifically through semantic segmentation techniques on satellite imagery, we streamlined the process of updating cadastral registries, reducing costs and time significantly. We utilized panoptic segmentation of building footprints from satellite images with a GSD resolution of 30cm/pixel upscaled to 15cm/pixel, and compared the updated building maps to old maps to detect changes. Our innovative approach gained substantial recognition within the Berkeley ecosystem, leading to our admission into Batch 15 of SkyDeck's Startup Incubation Program, winning the Berkeley Changemakers award, and initiating three significant projects. The complexities of working within an environment fraught with corruption and fraud ultimately led me to separate from the startup, but the experience underscored the enormous potential of AI-driven geospatial data analysis.

In my role as a project manager and tech lead, I developed a robust IoT-based full-lifecycle industrial analytics Platform as a Service (PaaS) specifically designed for the textile manufacturing industry in Mexico. This work presented significant technical challenges, from constructing a modular distributed network for machine digitalization using cutting-edge tools such as ESP32 development boards, Mongoose OS firmware, MQTT protocol, and AWS IoT Core, to engineering a system that could provide advanced analytics and monitoring capabilities like real-time data visualization and predictive maintenance. I implemented a complex data pipeline capable of handling over 753K machine state changes per day from 50 different nodes, transforming these messages into a DynamoDB to enable sophisticated machine learning-driven analytics. Furthermore, I developed a real-time analytics web application that delivered intuitive and interactive dashboards for data-driven decision-making. The entire project required seamless integration, high scalability, and precise quality control, and its successful deployment stands as a testament to the innovative engineering and rigorous problem-solving that characterized this experience.

In my tenure as a developer at Allie Cobranza Especializada, I executed the implementation of a cloud-based data analysis platform to optimize debt collection services, addressing an intricate problem that involved a diverse set of technical and analytical challenges. The project required the processing of a comprehensive dataset of 312.7K individual debtors and the employment of advanced analytical techniques, including geospatial analysis methods that enabled the generation of 24.9K accurate street addresses from partial information. Collaborating with cross-functional teams, I designed and implemented web-scraping and data collection software that led to an annual cost savings of $1.5 million and utilized Geographic Information Systems (GIS) to identify poorly registered homes, expanding the total addressable market by over 11%. Furthermore, I was instrumental in improving the collection success rate from 37% to 67% and expediting the payment data processing workflow by a factor of 7x. From developing targeted collection strategies to deploying web-scraping tools, the complexity of the tasks undertaken not only resulted in significant improvements in the company's efficiency but also allowed me to provide valuable insights and recommendations directly to the CTO and government officials, influencing decision-making processes and contributing to broader performance enhancements.

In the Spring of 2023, as an Academic Student Employee for EECS 16B, I led discussions on designing modern information devices and systems, focusing on machine learning, circuit design, control, and signal processing for approximately 60 students per week. I also provided guidance to a class of 417 students, assisting with homework and exam preparation. This role required not only a deep understanding of complex technical concepts but also the ability to communicate these ideas effectively, fostering the growth and development of budding engineers and researchers.

My AI-powered semantic search engine for course content
Yi Ma's breakdown of the class

Based on Job Barron's website.