Lightnews — Scholar-powered news

Thiago Serra

@thserra.bsky.social

3.1K followers 1.1K following 580 posts

Assistant professor at University of Iowa, formerly at Bucknell University, mathematical optimizer with an #orms PhD from Carnegie Mellon University, curious about scaling up constraint learning, proud father of two

Posts Media Videos Starter Packs

Pinned

Thiago Serra @thserra.bsky.social · May 6

We have a new paper out, with some exciting results on how to solve optimization models embedding neural networks!

📝 Optimization over Trained (and Sparse) Neural Networks: A Surrogate within a Surrogate

🔗 arxiv.org/abs/2505.01985

🎤 Talk in Montreal tomorrow: www.crmath.ca/en/activitie...

1/N

Optimization over Trained (and Sparse) Neural Networks: A Surrogate within a Surrogate

We can approximate a constraint or an objective function that is uncertain or nonlinear with a neural network that we embed in the optimization model. This approach, which is known as constraint learn...

arxiv.org

1 11

Thiago Serra @thserra.bsky.social · 1d

🔥🔥🔥

Thiago Serra @thserra.bsky.social · 2d

Last, but not least, I also managed to see how fast my little niece is growing during my quick hop connecting in Sao Paulo! 😍

3/3

Thiago Serra @thserra.bsky.social · 2d

This was a great opportunity for strengthening my ties with the Brazilian AI community and meet many talented colleagues. I am very grateful for the invitation by Rosiane de Freitas-Rodrigues to play such an important role at the conference, and for reconnecting me with my academic origins.

2/3

Thiago Serra @thserra.bsky.social · 2d

Last week I had the immense privilege of giving the opening keynote talk of the BRACIS - Brazilian Conference on Intelligent Systems #bracis2025 and visiting the beautiful city of Fortaleza for the first time.

1/3

1 2

Thiago Serra @thserra.bsky.social · 4d

Second, she has partnered with SAS Brasil to first train people who will produce a better and more inclusive melanoma dataset. With a sensitization of the human who diagnoses about what types of lesions are malign on different skin tones, she hopes that we will eventually train the right model!

5/N

Thiago Serra @thserra.bsky.social · 4d

The recognition of her work led to new and unexpected references and collaborations. First, the cosmetic industry is considerably ahead in cataloging skin tones, and in fact celebrated her contributions.

4/N

Thiago Serra @thserra.bsky.social · 4d

On the other hand, those models missed the mark in lesions at body extremities like the hand palm, foot sole, and under the nails; which have considerably higher incidence in darker-skinned individuals.

3/N

Thiago Serra @thserra.bsky.social · 4d

By removing the skin lesions entirely from the images in one way or another, her team still obtained models that were still considerably accurate due to other characteristics present in those images.

2/N

Thiago Serra @thserra.bsky.social · 4d

Closing #bracis2025 with mastery, Sandra Avila talked about her journey working on diagnosing melanomas (a type of skin cancer).

Interestingly, the story starts with winning a classification contest, and then evolves from her working backwards to figure out what the models actually predicted.

1/N

1 2

Thiago Serra @thserra.bsky.social · 7d

Although such attacks can be successful without interfering directly with the node being attacked, they lead to high-rank tensors. Hence, low-rank tensor approximations end up being a way to filtering the attacker’s noise from the information encoded in the graph.

6/6

Thiago Serra @thserra.bsky.social · 7d

Towards the end, Papalexakis talked about computationally efficient clustering methods, and how they may be bound to adversarial attacks in graph neural networks.

5/N

Thiago Serra @thserra.bsky.social · 7d

Rank-one tensors are blocks obtained as the product of vectors, which approximate part of the information in the tensor.

That allows to investigate networks in a number of ways, such as in one of their studies on citations, coauthorship, and clustering of scholars in different areas.

4/N

Thiago Serra @thserra.bsky.social · 7d

Papalexakis then uses the example of a graph in which multiple edges between the same vertices are possible, which leads to one adjacency matrix for each label and therefore to a “3-dimensional matrix”, or tensor, if we want all of those adjacency matrices as part of a same mathematical object.

3/N

Thiago Serra @thserra.bsky.social · 7d

Preparing the ground for a future generalization from matrices to tensors, Papalexakis characterized rank-one matrices as those that can be obtained from the product of a column and a row. That entails a perspective of low-rank approximations as a sum of few such column-row products.

2/N

Thiago Serra @thserra.bsky.social · 7d

Evangelos Papalexakis gave an interesting and accessible keynote about latent structure and tensor decomposition at #bracis2025

The identification of latent structure was motivated with the observation from Miller’s Law that we only keep “7 +/- 2” items in memory at any given time time.

1/N