João Pedro Pires T.

I study how to anchor decision boundaries to the intrinsic geometry of the data manifold.

Standard training lets decision boundaries cut blindly through the empty, low-density space between samples, leaving models fragile to adversarial perturbations and overconfident off-manifold.

I show that making the loss invariant to random, isotropic perturbations of the input is mathematically equivalent to regularizing the loss gradient with the score of the data distribution, via Stein's Identity.

This ties the learning process directly to the data manifold, forcing the network to rely only on gradients that are geometrically meaningful, rather than ones shaped by noise in empty ambient space.

Learn more

I study optimization and fine-tuning techniques for efficient LoRA and learning transfer.

Because MOGR regularizes the loss with the score of the data distribution, it shapes decision boundaries that are more effective and geometrically meaningful earlier in training.

On modular division tasks, score-regularized models reach high validation accuracy in a fraction of the optimization steps needed by standard training, closely related to the grokking phenomenon.

I'm extending this to LoRA and other parameter-efficient fine-tuning setups, where faster, more reliable convergence directly cuts the cost of adapting large models.

Learn more

I study how to detect and infer anomalous data in the context of likelihood estimation.

Normalizing Flows (NFs) are a powerful class of generative models.

Its ability to directly compute likelihood allow for a stable and efficient training.

Using deep architectures we can model the data distribution. Apart from generating new samples, the likelihood of samples are easily obtained.

Learn more

As part of my Petrobras research, I study and develop automatic seismic interpretation tools.

Faults are closely linked to hydrocarbon transport. They can be used for geological structural modeling, well location screening, and tectonic analysis.

Given the complex and challenging nature of seismic data, DL architectures are being used to automate seismic interpretation.

I explored a diverse array of architectures to address the problem of fault detection. An upcoming benchmark article with our results will soon be available!

Learn more

I studied how to automatically fit isochrones to a star cluster in a colour-magnitude diagram.

Isochorones are theoretical curves used to estimate a stellar population's age, distance, reddening, and metalicity.

Although very reliable, these curves are mainly fitted manually. This makes impractical to study thousands of clusters, not to mention author divergence.

Using parallelism, I made a CPU-bound algorithm that outruns existing MCMC routines.

Learn more

I studied the relationship between open star clusters with MW's substructures, such as stellar streams.

Open clusters are located in the Milky Way's disk. In fact, they constitute our Galaxy spiral arms.

With recent census, data showed that some open clusters are located in the thin disk and also in the halo.

I studied how these clusters ended up in these regions of Galaxy and tried to identify possible links between the clusters orbit and merger or flyby events.

Learn more

I applied unsupervised clustering algorithms to identify and charcterize open star clusters.

With the Gaia revolution, many stellar clusters were discovered and many more are yet to be discovered.

However, the star-members list among authors can present as high as 20% divergences.

Using unsupervised clustering algorithms, I studied how to automate the process of identifying and characterizing open star clusters.

Learn more