Gabriel Gutierrez
ML ResearcherSoftware ArchitectOpen Source
Deep Learning,
Software Engineering
& Computer Vision
Meet Gabriel
About Me
Machine Learning
BERTNLPSeismic FaciesWeak SupervisionUnicampEAGE
I developed a passion for machine learning during my university studies, where I explored its vast potential to solve real-world problems. This interest led me to work as a junior researcher volunteer for two years, focusing on natural language processing, specifically fake news detection using BERT.I completed my master's degree at Unicamp with a dissertation on the semantic segmentation of seismic facies using transformer architectures, with findings published in Geophysical Prospecting and presented at the 85th annual EAGE conference. Building on that foundation, I am currently pursuing a PhD at Unicamp, where my research focuses on weakly supervised semantic segmentation — developing methods that learn from incomplete, noisy, or inaccurate annotations rather than requiring exhaustive pixel-level labels.
Software Engineering
Open SourceMinervaArchitecturePyTorch LightningDiscovery Unicamp
I hold a degree in Software Engineering from one of Brazil's top universities, where I contributed to projects focused on medical data collection for research aimed at informing public policy decisions.Since joining Unicamp's graduate program, I have applied my software engineering expertise directly to research, bringing a rigorous engineering perspective to complex scientific challenges. I lead Minerva, an open-source machine learning training framework developed in collaboration with Discovery Unicamp. As software architect, core developer, and core maintainer, I design the framework to be robust, modular, and aligned with the evolving demands of the research community — integrating tools such as PyTorch Lightning, Ray Tune, Hyperopt, and MMsegmentation.
My Skills
Overview
Computer VisionSemantic SegmentationML FrameworksOpen SourceResearch
With a strong foundation in software engineering and machine learning, I bring a unique blend of technical expertise and research experience to my projects. My background spans developing data-driven applications, contributing to impactful research, and leading open-source initiatives. I specialize in computer vision, semantic segmentation, and building scalable machine learning frameworks like Minerva. My skills extend beyond coding; I architect solutions that bridge the gap between cutting-edge research and practical implementation, always with a focus on innovation and collaboration.Transformers
BERTViTMiTSegFormerSETRFoundation Models
I have extensive experience working with transformer models, beginning with BERT during my early research on fake news detection. In my master's program I expanded this to vision transformers — working with ViT, MiT, SegFormer, Segmenter, and SETR-PUP for seismic facies segmentation. In my PhD, I continue to build on transformer-based architectures, exploring how large pretrained models and foundation models can be leveraged under weak supervision, where only coarse or noisy labels are available.Computer Vision
SegmentationWeak SupervisionScientific Imaging
My experience in computer vision spans semantic segmentation, image processing, and weakly supervised learning. In my master's research I worked extensively with transformer and CNN-based segmentation models applied to seismic data. My PhD extends this toward label-efficient learning, where I investigate how models can be trained effectively under noisy or incomplete supervision. I have developed a deep understanding of the challenges specific to scientific imaging domains, where annotation is expensive and ground truth is inherently uncertain.Weak Supervision
Pseudo-labelsSAMNoisy AnnotationsFoundation Models
My current research centers on weakly supervised semantic segmentation — a paradigm where models learn from imprecise, incomplete, or noisy annotations rather than costly pixel-level labels. This includes working with inaccurate supervision, pseudo-label generation and correction, and SAM-based mask refinement. I study how foundation models and self-supervised pretraining can compensate for annotation scarcity, combining insights from the survey literature with exploration of recent advances in the field.My Work
Semantic Segmentation
DeepLabSegFormerSETRF3 DatasetSeam AIGeophysical Prospecting
In my master's dissertation at Unicamp, I conducted a systematic evaluation of transformer-based segmentation models applied to seismic facies analysis. The study compared five architectures — DeepLab V3, DeepLab V3+, SegFormer, Segmenter, and SETR — trained and evaluated under identical conditions on the F3 and Seam AI datasets, directly addressing the reproducibility gap prevalent in the field. We found that SETR shows promising performance on both datasets, while CNN models offer a higher performance-to-parameter-count ratio compared to transformer models.This work was published in Geophysical Prospecting under the title 'On the Performance Evaluation of Deep Learning Models for Seismic Facies Segmentation', and was also presented at the 85th annual EAGE conference and exhibition, where both the paper and the presentation received among the highest ratings from reviewers.
PhD Research
Weakly Supervised Semantic Segmentation
Weak SupervisionSAMPseudo-labelsNoisy LabelsUnicamp
My PhD research at Unicamp investigates weakly supervised semantic segmentation — the problem of training dense prediction models without exhaustive pixel-level annotation. Specifically, I focus on inaccurate and noisy label supervision, where training masks are available but contain errors, imprecisions, or incomplete coverage. This setting is particularly relevant to scientific domains such as geoscience, where expert annotation is expensive and inherently subjective.My work draws on recent advances in pseudo-label refinement, Segment Anything Model (SAM) based mask correction, and foundation model adaptation, combining rigorous software engineering with applied research to develop methods that are both principled and practically deployable. This line of research builds directly on the seismic facies segmentation work from my master's, extending it toward more realistic, label-efficient scenarios.
Minerva
View on GitHubOpen SourceML FrameworkPythonDiscovery Unicamp
As the lead of Minerva, an open-source machine learning training framework developed at Discovery Unicamp, I've played a key role in its architecture, development, and maintenance. Minerva was designed to meet the specific needs of researchers, providing a flexible and scalable solution for training and fine-tuning machine learning models across a variety of tasks. My role as the software architect, core developer, and core maintainer has involved ensuring that the framework is robust, user-friendly, and adaptable to the fast-evolving demands of machine learning research.Minerva stands out by offering an accessible and powerful tool for both academic and industry researchers, bridging the gap between research and practical implementation. Through my work on this project, I've gained valuable experience in open-source software development, contributing to a tool that empowers the research community to advance machine learning in meaningful ways.