Luis Leal

:) arxiv.org/abs/2010.10004

1987 ⚡️🔥 Keeper of the Seven Keys: Part I forever 🤘 Which song is untouchable for you? 👇 #Helloween

CEO's learning Ai will cost more than human workers after firing 25% of their workforce

We used to go to a special website, ask strangers for help with programming, and get humiliated in return

One theorem every ML engineer should know: The Bellman Optimality Principle. It states that the optimal solution to a decision problem can be constructed recursively from optimal subproblems. In reinforcement learning, this becomes: Why it matters: • Foundation of Q-learning and dynamic programming • Enables sequential decision-making under uncertainty • Central to robotics, game AI, and autonomous systems • Connects optimization with learning The profound idea: Intelligence can emerge from recursively improving future decisions. Almost every modern RL algorithm — from DQN to AlphaGo — builds on Bellman’s insight. Reinforcement learning is ultimately the mathematics of long-term consequences. Image: share.google/AIBaxXi8u61KVl…

Linear Algebra and Optimization for Machine Learning [516-page textbook]: amzn.to/39aWf8N ————— #DataScience #DataScientist #ML #Mathematics #ORMS

-Soy un Nazgül, uno de los Jinetes Negros, Espectro del Anillo y servidor de Sauron, el Señor Oscuro… - ¿Ok, y el perrito? - Ah, él es Firulais, pero no tenía con quien dejarlo y no le gusta quedarse solito…

Taylor Series aren’t just a calculus topic, they’re foundational to ML optimization. My latest video with @Cohere_Labs breaks down how local approximations of loss functions lead to Gradient Descent and Newton’s Method. Watch here: youtu.be/ALP1J81Sqxw #MachineLearning #AI

Valiste madre, hantavirus 💀…

One theorem every ML engineer should know: The Universal Approximation Theorem. A neural network with even a single hidden layer can approximate almost any continuous function — given enough neurons. Why it matters: • Explains why neural networks are so expressive • Connects deep learning with function approximation • Forms a theoretical foundation for modern AI But here’s the catch: The theorem guarantees existence, not efficient training. A network can represent the function. That doesn’t mean gradient descent will find it. Deep learning is as much about optimization as approximation.

Hermite Expansion is a powerful mathematical tool used to represent functions through Hermite polynomials, which form an orthogonal basis under Gaussian measures. Originally developed in probability theory and mathematical physics, Hermite expansions later became important in stochastic analysis, approximation theory, and functional analysis. In machine learning (ML), Hermite expansions are useful because many models involve Gaussian assumptions in data, noise, or latent variables. They appear in kernel methods, random feature models, and nonlinear approximation techniques. Hermite-based representations often help simplify high-dimensional learning problems and provide theoretical insight into model behavior. In deep learning (DL), Hermite methods are used to analyze neural network dynamics, activation functions, and representation learning. Researchers use them to study how neural networks approximate complex nonlinear functions and how information propagates through deep architectures. In reinforcement learning (RL), Hermite expansions appear in stochastic control, value function approximation, and diffusion-based decision models. They are particularly useful in noisy environments where uncertainty and Gaussian processes play central roles. Hermite techniques also connect with Wiener chaos expansions, helping researchers analyze stochastic gradients, uncertainty propagation, and learning dynamics in modern AI systems. Image: share.google/ebKboTCZiRiU8D…

Qué gran época fue la del maestro Odiomistweets 🐦‍⬛

No se peleen en el tráfico por vida suyaaaa! Dejen que se vaya ese necio, que les tira el carro, que anda bocinando como loco etc. Que se mate solo, pero no se metan en problemas por una cosa de esas. Su vida vale más que eso.

Según la UNESCO, registra que tenemos 1 investigador dedicado a investigación a tiempo completo por cada millón de hábitantes. Desolador.

😀

Discúlpenme.

offline RL con CQL(conservative q-learning) + SEM(simplicial embeddings)

CFOs realizing that their AI token budget is going to be higher than the salaries of the people they laid off