Lucas Porto Maziero

Explore José Antonio Siqueira de Cerqueira's efforts in enhancing language models for greater safety and reliability. Dive into this article to discover his significant contributions to the field - https://lnkd.in/dA-EWMu2! 📚👨‍💻 #research #languagemodels

4

We are excited to announce 3rd Workshop on Enabling Software Engineering Practices via latest development trends (ESEP 2025) held in Florianópolis, at Federal University of Santa Catarina, Brazil, 15 - 17 April 2025. This workshop is a part of WorldCIST'25 - 13rd World Conference on Information Systems and Technologies. This workshop aims to provide a meeting point for both practitioners and researchers to discuss the theoretical and empirical studies conducted to evolve and manage Software Engineering (SE) practices and resolve related issues by employing modern technology. The ever-evolving nature of the software development industry due to the inclusion of new development trends such as Machine Learning (ML) operations, DevSecOps, Serverless Computing, Blockchain, Internet of Things (IoT), Big Data, Augmented Reality (AR) and quantum computing lead to introduce new software engineering practices. This workshop will provide an opportunity to present empirical evidence of how the latest software development trends are enabling software engineering practices addressing the issue in the context of effort estimation, software process improvement, quality assessment, risk management, continuous integration, and so on. Topics of interest include (but are not limited to): Blockchain User Experience Design Machine Learning Operations Vulnerability Disclosure Programs Continuous Integration and Delivery Quantum Computing Edge Computing Serverless Computing DevSecOps Augmented Reality (AR) Digitalization of software process and products Use of social networks Knowledge management from semi and unstructured data Data-driven modeling Data-driven approaches for risk management Software development issues and challenges due to emerging trends Data and Business Analytics https://lnkd.in/dzV-egBs Submission: Dec. 1, 2024 Notification: Jan. 5, 2025 Registration: Jan. 15, 2025 #Springer #Research #SoftwareEngineering #Workshop #Conference #2025

12

Vitor Cerqueira presented our work at Discovery Science 2024 in Pisa, Italy! 🌍 Our paper, "Forecasting with Deep Learning: Beyond Average of Average of Average Performance," introduces a new Aspect-based Evaluation Framework that moves beyond the traditional focus on single summary metrics. This work is also co-authored by Carlos Soares. Our approach evaluates forecasting models by analyzing performance across several key dimensions: sampling frequency, forecasting horizon, problem difficulty, and anomalies (extreme cases outside the 99% prediction interval). To provide a clearer understanding of model strengths and weaknesses, we use different summarization techniques, including overall performance metrics, expected shortfall (focusing on how models perform in the worst 5% of cases), and win/loss ratios for comparing models without assuming a specific distribution. We also incorporate the region of practical equivalence (ROPE) to focus on differences that are practically significant. We showed how controlling for these dimensions gives a clearer, more reliable view of model performance. This approach leads to more informed model selection and a better understanding of trade-offs, which is crucial for responsible forecasting. Link to our paper: https://lnkd.in/d3EyUBHb #DiscoveryScience2024 #Forecasting #DeepLearning #AI #Research #TimeSeries #ModelEvaluation #ResponsibleAI #DataScience

128

2 comentários

An interesting talk about intersection between mathematics and computer science (to help on proving theorems). Lean is mentioned as Terence Tao is a heavy user of this programming language developed by a Brazilian guy named Leonardo de Moura. Thanks, Dirceu Barquette Filho, for the recommendation. #mathematics #theoremproving #lean #computerscience https://lnkd.in/daXYfJnG

15

3 comentários

📚 Just came across an fascinating set of lecture notes on "𝐓𝐞𝐧 𝐋𝐞𝐜𝐭𝐮𝐫𝐞𝐬 𝐚𝐧𝐝 𝐅𝐨𝐫𝐭𝐲-𝐓𝐰𝐨 𝐎𝐩𝐞𝐧 𝐏𝐫𝐨𝐛𝐥𝐞𝐦𝐬 𝐢𝐧 𝐭𝐡𝐞 𝐌𝐚𝐭𝐡𝐞𝐦𝐚𝐭𝐢𝐜𝐬 𝐨𝐟 𝐃𝐚𝐭𝐚 𝐒𝐜𝐢𝐞𝐧𝐜𝐞" by Afonso S. Bandeira. 🎩 This comprehensive work presents 42 open problems at the intersection of mathematics and modern data analysis. 🔑 Key highlights: - Deep dives into Principal Component Analysis in high dimensions, including critical insights on when PCA works (and when it doesn't) - Novel perspectives on spectral clustering and graph-based learning methods - Rigorous treatment of concentration inequalities and their applications to random matrices - Cutting-edge material on compressed sensing and sparse recovery - Elegant connections between seemingly disparate areas like coding theory and community detection 👀 What I found most compelling was how the notes bridge pure mathematical concepts with practical data science applications. The author does an excellent job explaining how theoretical results translate to real-world algorithmic improvements. 🧵 The open problems presented range from fundamental questions in random matrix theory to practical challenges in data clustering and signal processing. Many remain unsolved and could lead to breakthrough advances in how we process and analyze data. 🧵 If you're interested in the mathematical foundations of data science, this is definitely worth checking out. It's particularly valuable for those working at the intersection of mathematics, computer science, and data analysis. #Mathematics #DataScience #MachineLearning #Research #OpenProblems 👇👇👇 Im creating a lot of scientific content. You're invited to follow me at: 🎙️ Spotify: - https://lnkd.in/d2Y-ehe7 - https://lnkd.in/d-gMtCrE (Hebrew) 🚇 Youtube: https://lnkd.in/dQNqdvMS ☎️ Telegram: https://lnkd.in/dVVqhNw5 (in Hebrew) 🦜 Twitter: https://lnkd.in/dTse8avN

229

6 comentários

Check out my latest Medium article! We dive into two different approaches to fine-tune the Donut Transformer for document classification, optimizing efficiency and accuracy in health insurance claims. I hope you enjoy it!

27

2 comentários

I took a graduate database course at the University of São Paulo 13 years ago where the professor said: "NoSQL databases are all the rage and we'll cover them at the end of the course. But first you need to understand relational databases to know everything you loose when you choose NoSQL: everything the NoSQL database doesn't do for you which you'll have to reinvent, implement, and test in your application. Essential features like transactions updating multiple records reliably and safely in the face of concurrent requests." If you're not at the scale of thousands of servers, it's most likely that a relational database will make your life EASIER and your app MORE RELIABLE than a "cool" NoSQL alternative.

528

42 comentários

📊 Exciting news! 📊 Sharing some valuable data with the community on my project Kindala! 🚀 Check out KindalaOpenData, my public repository, for historical price data from May to June 2024 for stores in Brasilia, DF. 🏬 More data and code coming soon! 💻 Utilizing this data for time series analysis - possibilities are endless! 📈 Data is under The MIT License, so feel free to use it however you'd like, just remember to credit the Kindala Project. 🙌 All the Best.. #DataAnalysis #TimeSeries #OpenData #KindalaProject

6

Thanks Valentina Alto for share your knowledge! Some of the topics discussed included: - Fundamentals of prompting and how it powers LLM applications - Principles and advanced techniques for strategic prompt engineering - Techniques to reduce risks like hallucination through careful prompting - Approaches to chain prompting for multi-step reasoning - Defensive techniques to secure prompts against injection attacks Ways to protect prompts against hacking attempts and misuse #LLMPromptStrategies #GenAI #PromptEngineering

13

In C++, there are different ways to pass a value to a function. Typically, at any given time, an object in C++ ‘belongs’ to a single function. The various ways to call a function differ in who owns the object, the caller or the callee (the function being called). The simplest one is that we pass by value. In such cases, a copy is typically made of the object and both the caller and the callee own a copy. void value(MyClass obj) {} We can pass by reference. You recognize a reference by the single ampersand (&). The caller owns the object, but the callee gets access to it. void reference(MyClass& obj) {} You can also pass by an “rvalue reference” which you recognize by the ‘&&’ symbols. In such cases while the caller initially creates the object, but its ownership is passed to the callee. I personally dislike the expression ‘rvalue reference’ and I would have preferred something less technical. void rvalue_reference(MyClass&& obj) {} However, in some instances, you do not care whether your function gets to own the value, or has merely a reference to it. Writing two functions duplicates the code. Instead, you can then use a forwarding reference: template <typename T> void forwarding_reference(T&& obj) {} It looks like an rvalue reference, but it is not: it can be either a normal reference or an rvalue reference depending on how you call it. Here is how you might call these functions in practice: MyClass obj; value(obj); reference(obj); rvalue_reference(MyClass()); forwarding_reference(obj); forwarding_reference(MyClass()); https://lnkd.in/eMR-2cTw

50

3 comentários

"Mastering SVM: A Guide to Different Kernel Functions 🧠✨ Support Vector Machines (SVM) are versatile tools in the machine learning toolkit, but understanding how to choose the right kernel function is key to unlocking their full potential 🔓. Here’s a deeper dive into the most popular SVM kernels: 📈 Linear Kernel: Best for when your data can be separated with a straight line 📊. This kernel is simple and efficient, making it perfect for high-dimensional datasets where a clear boundary exists 🔒. 📊 Polynomial Kernel: This kernel can capture more complex relationships in your data by considering polynomial combinations of features 🔍. It’s ideal when your data isn't linearly separable but still follows some predictable pattern 📈. ⚡️ Radial Basis Function (RBF) Kernel: Known as the Gaussian kernel, it's the most flexible choice 🤯. The RBF kernel can handle complex, non-linear relationships by mapping the data into a higher-dimensional space, making it powerful for real-world, messy datasets 🌐. Choosing the right kernel function is crucial to maximizing your SVM model’s accuracy and effectiveness 📊. Whether you need simplicity or the ability to model complex data relationships, there’s a kernel function for every need 🎯! Dive deeper into SVMs and elevate your machine learning game today! 🚀 #MachineLearning #SVM #KernelFunctions #DataScience #ArtificialIntelligence #TechTalk #DataAnalytics #AI 🤖"#SkillzRevo 💻 #NorthAmerica 🇺🇸 #USA 🇺🇸 #Dubai 🇦🇪 #MiddleEast 🌍"

4

I'm happy to share that our team (Emilio Vital Brazil, Victor Yukio Shirasuna, Renato Cerqueira, Dmitry Zubarev, Kristin Schmidt, Indra Priyadarsini, Seiji Takeda, Nathaniel Park) has had five papers accepted at NeurIPS in the AI for Materials and Foundation Models for Science workshops. We'll be in City of Vancouver from Dec 10 to Dec 15 to present our work and connect with the community. The papers we’ll be presenting are: 1. "A Mamba-Based Foundation Model for Chemistry"   2. "A Large Encoder-Decoder Polymer-Based Foundation Model"   3. "Multi-View Mixture-of-Experts for Predicting Molecular Properties Using SMILES, SELFIES, and Graph-Based Representations"   4. "SELF-BART: A Transformer-based Molecular Representation Model using SELFIES"   5. "Agnostic Causality-Driven Enhancement of Chemical Foundation Models on Downstream Tasks" All pre-prints will be available soon. Additionally, part of this work will be open-sourced as part of IBM's Foundation Models for Materials initiative: (https://lnkd.in/du-_ET8H).

83

5 comentários

Paper summary: GQA: Training Generalised Multi-Query Transformer Models from Multi-Head Checkpoints Why I care: Used in the Llama 3 paper, GQA can be applied to any model using Multi-Head Attention to speed up inference. Link: https://lnkd.in/eEhiFGBq Key points: • Multi-Query Attention (MQA) is a variant of Multi-Head Attention (MHA), where, instead of each attention head having its own set of keys, values, and queries (as in MHA), each head in MQA shares the same keys and values, but has a different set of queries. This design reduces memory and computation requirements during inference, trading off some model expressiveness for a significant boost in speed, at the cost of a slight decrease in performance. Additionally, models using MQA often require retraining or fine-tuning to achieve optimal performance. • This paper introduces two improvements to standard MQA. First, it presents a method for uptraining MHA models into MQA models using only 5% of the original pre-training compute. Second, it introduces a generalization of MQA, called Grouped-Query Attention (GQA), which achieves quality similar to MHA with speeds comparable to MQA. • To transform an MHA model into an MQA, the authors propose a two-step process. First, the checkpoint is converted to the new format, with the best solution they tested involving mean pooling the projection matrices for the key and value heads. Second, the checkpoint is pre-trained for a small fraction of the original training steps, e.g., 5%. • GQA-G is an interpolation between MQA (GQA-1) and MHA (GQA-H), where instead of having just one key and value head or having all key and value heads, this is treated as a hyperparameter that can be selected. For example, if G = H/2, this means that for every key and value head, there are two query heads. • MQA reduces the size of the key-value cache by a factor of H, representing an aggressive cut in both memory bandwidth and capacity. Larger models suffer relatively less from memory bandwidth overhead from attention, as the KV cache scales with model dimension, while FLOPs and parameters scale with the square of the model dimension. Additionally, standard sharding for large models replicates the single key and value head across the number of model partitions. GQA eliminates the waste from such partitioning and is particularly a good trade-off for larger models. • Empirically, the authors found that GQA-8 offers the best trade-off between performance and speed, with 8 being the same value used by Llama 3. Follow me on X (Twitter) for more: https://x.com/mnslarcher

15

Yesterday, I shared a brief tutorial on deriving digital terrain models using only a digital surface model in plimanshiny. I was curious about the accuracy of this approach, so I decided to conduct a quick evaluation. Using an example from Filipe Matias's BisonFly Pipeline (https://lnkd.in/d2-3VJiu), I constructed the canopy height model using two methods: 1) Original DSM - Original DTM, and 2) Original DSM - Derived DTM from DSM. This can now be implemented using pliman::mosaic_chm(dsm), where dsm is the digital surface model. The DTM derived from the DSM using a moving window approach resulted in an estimated digital terrain model with a mean error of 0.001758105 meters compared to the original ones. Consequently, both volume and plant height estimated using the derived DTM were highly correlated with the "true" values. This new approach can accelerate the high-throughput phenotyping pipeline because it eliminates the need for a prior flight before sowing or sampling points for digital surface models, in addition to reducing the between-flight error, since the digital terrain model is derived within each flight. I foresee the possibility of implementing this in plimanshiny's time series module to compute, for example, a growth model with plant height or volume, which could be incredibly beneficial for many research areas.

140

1 comentário

Ever wanted to train your own octopus detector? ️🤖👁️🐙 In this getting started guide for Chainguard's secure image for #PyTorch and #CUDA, we step through training a computer vision model to detect octopuses 🐙, whales 🐋, and penguins 🐧. (Representing Chainguard's Wolfi OS, #Docker, and Linux, if you're following along at home.) And, yes, the PyTorch image has Chainguard's trademark 0 CVEs (vulnerabilities) as of today. This is a great place to get started with deploying #AI training and inference securely. And you'll be able to detect those sneaky octopuses before they get the jump on you... We're in beta for these images, so drop us a line if you'd like to be part of the program. https://lnkd.in/etMWceuE

21

One problem from new version of gfortran. When the new version of mpi compiler with gfortran is used to compile the code, a new problem comes out. The dummy variable for some subroutines from MPI gets a problem, e.g., it can be a real number in one call, and it can be an integer in another call. This brings a problem with gfortran compiler. An option can be avoided with a switch from compiler, -fallow-argument-mismatch.

1

Excited to share that at Deeplite, we’ve recently published our research at ECCV 2024. 🎉 Our paper benchmarks 131 tiny YOLO models optimized for resource-constrained hardware like MCUs. You can check it out here: https://lnkd.in/dJNS2CSu. We’ve re-engineered the necks and backbones of the native YOLO family models, fine-tuning over 1000 architectures across diverse datasets like Microsoft COCO and VOC. Each architecture was evaluated based on key trade-offs, including accuracy vs. latency, accuracy vs. RAM, and Flash usage. Our paper presents the final rankings of all these models, offering valuable insights for selecting the right architecture based on specific MCU hardware constraints. Congratulations to the entire Deeplite team, and special mention to all the authors: Sudhakar Sah (Sud), Darshan C G, Matteo Grimaldi, Honnesh Rohmetra, and Ehsan Saboori (PhD, Eng). #research #ECCV2024 #deeplearning #AI #tinyYOLO #embeddedAI #computervision

50

6 comentários