Pedro C. Neto
@Pedro18_Neto
Followers
213
Following
3K
Media
98
Statuses
2K
Artificial Intelligence Scientist at Unilabs. PhD from FEUP. Invited Assistant Professor at FEUP. 🔙 Aalto University, Finland and ISEP, Portugal
Joined November 2015
We have recently proposed a novel take on xAI. Taking advantage of the duality between xAI and Causality, we have proposed a Causality-inspired taxonomy for explainable artificial intelligence. https://t.co/b2CSHZuGkR Follow the thread 🧵 to know more!
arxiv.org
As two sides of the same coin, causality and explainable artificial intelligence (xAI) were initially proposed and developed with different goals. However, the latter can only be complete when...
1
0
3
Finally finished my PhD. A long journey that comes to an end! Great defense, with very interesting discussion! #PhDone
1
0
2
Já está disponível a minha primeira opinião pública sobre o uso da IA na saúde, tanto em Portugal como no mundo!
Quão loucos somos para confiar a nossa saúde a sistemas que não foram desenhados a pensar nela, que carecem de consistência de anos de investigação que fundamentam o conhecimento humano especializado? Crónica de Pedro C. Neto
0
0
1
Quão loucos somos para confiar a nossa saúde a sistemas que não foram desenhados a pensar nela, que carecem de consistência de anos de investigação que fundamentam o conhecimento humano especializado? Crónica de Pedro C. Neto
publico.pt
Quão loucos somos para confiar a nossa saúde a sistemas que não foram desenhados a pensar nela, que carecem de consistência de anos de investigação que fundamentam o conhecimento humano especializado?
0
1
1
Many of you do not know the trauma of having to write out Java on paper for the comp sci AP exam and it shows
314
1K
18K
This is the last paper of my PhD, and the one I care about most. #AI #FaceRecognition #Fairness #BiasInAI #ComputerVision #PhD #DeepLearning #EthicalAI #RepresentationMatters #arXiv
0
0
1
🧠 We hope this sparks new ways to: → Think about fairness → Annotate data → Evaluate models → Build inclusive systems 📄 Full paper: https://t.co/FfxxW58PLW 🙌 Feedback, shares & collabs welcome! 🧵 7/8
arxiv.org
Bias has been a constant in face recognition models. Over the years, researchers have looked at it from both the model and the data point of view. However, their approach to mitigation of data...
1
0
0
One of my favorite results: ✨Fair AI doesn’t mean pushing everyone into a few fixed categories. It means modeling the real diversity of people. 🧵 6/8
1
0
0
We also show something crucial: ❌ Equal samples per group ≠ Fair representation. Most benchmarks balance identities across groups like “10K Asian, 10K Black, 10K White.” But real-world data is messier — and imbalanced. Our method works with this complexity. 🧵 5/8
1
0
0
🚀 The results: → Continuous-label training consistently outperforms traditional approaches. → Fairness improves significantly — especially for underrepresented groups. → And no, we don’t sacrifice accuracy. We get both fairness + performance. 🧵 4/8
1
0
0
📌 In our paper, we: ✔️ Model ethnicity as a continuous label (based on human-perceived similarity) ✔️ Propose a distribution-aware sampling strategy ✔️ Train 65+ models on 20+ dataset variations The goal? Make AI less biased and more representative 🧵 3/8
1
0
0
👤 Most face recognition fairness studies group people into discrete categories: “White”, “Asian”, “Black”, etc. But what if that’s the wrong starting point? ➡️ People don’t fit into boxes. Ethnicity is complex — and often continuous. 🧵 2/8
1
0
0
🚨 Just out: the final paper of my PhD! We challenge the way face recognition “does fairness.” 🔍 What if the problem isn't just bias — but the way we label humans? 👇 A thread on continuous demographic labels and why they matter. 📄 https://t.co/ByGpH73eOv 🧵 1/8
arxiv.org
Bias has been a constant in face recognition models. Over the years, researchers have looked at it from both the model and the data point of view. However, their approach to mitigation of data...
1
0
0
All the energy missing during the blackout has been released today and last night from the skies… what a thunderstorm
0
0
0
One just needs Takamura
0
0
0
Just trained ArcFace on both M1 Pro and M3 Max, using PyTorch and mps… the M3 Max is 100% faster, leading to half the training time! Yet far from a Nvidia A100 (as expected) but truly fast! Nice one @Apple
0
0
0
@Aidamo27 Great work @Aidamo27 ! We also observe behavioral differences in quantized models, with increased values on bias-related metrics! In some cases mixing synthetic samples helped. https://t.co/hvvReJ6tVx
https://t.co/0nb3a4HTEm
ieeexplore.ieee.org
With the ever-growing complexity of deep learning models for face recognition, it becomes hard to deploy these systems in real life. Researchers have two options: 1) use smaller models; 2) compress...
0
1
1
@Aidamo27 We studied the possibility of distilling from highly biased teachers (each specialist in a specific demographic group), and it seems to help with bias mitigation during the distillation process https://t.co/1JRTtEqcDV
0
1
2