Machine Learning Engineer: 20th May 2025

Active Learning vs. Data Filtering: The Key Difference (blog.blackhc.net, 2025-05-16). Active learning and data filtering differ in data selection timing and informativeness scores, with submodularity affecting how samples are chosen. Tools like BALD and BatchBALD aid in these processes

Multi-Omics Integration Strategy and Deep Diving into MOFA2 (divingintogeneticsandgenomics.com, 2025-05-17). Explore multi-omics integration strategies using MOFA2, a Bayesian probabilistic framework that identifies latent factors across omics layers, accommodating data sparsity while ensuring biological interpretability and validation

Spatial Resampling in Predictive Modelling (jmsallan.netlify.app, 2025-05-15). Explore spatial resampling in predictive modeling using tidymodels and spatialsample, demonstrated through the cat_adoption dataset to predict rescue cat outcomes based on geographic attributes

One tree is not enough – Random forest (weeklycspaper.com, 2025-05-13). Combining multiple decision trees into a random forest improves generalization and reduces bias, leveraging random sampling of features and robust performance against noise, with implications for both classification and regression tasks

Strength in Numbers: Ensembling Models with Bagging and Boosting (towardsdatascience.com, 2025-05-15). Bagging and boosting are essential ensemble techniques in machine learning, as they reduce variance and bias respectively, utilizing algorithms like decision trees to improve model performance on datasets

Climbing trees 1: what are decision trees? (mathpn.com, 2025-05-18). Decision trees are foundational in machine learning, utilized for both classification and regression tasks. Key algorithms include ID3, C4.5, and CART, focusing on splitting data to improve predictions

Building a Trading Strategy using Bias-Variance Decomposition (blog.quantinsti.com, 2025-05-13). Explore the bias-variance tradeoff in ML for trading strategies, utilizing PCA, VIF, and backtesting methods to optimize models and reduce overfitting in financial markets

Are Sector-Specific Machine Learning Models Better Than Generalists? (quantpedia.com, 2025-05-14). Research shows that Hybrid machine learning models, which combine industry awareness with broad training data, outperform Specialist and Generalist models in predicting stock returns, yielding higher Sharpe ratios and lower portfolio volatility

Fine-Tuning Transaction User Models (building.nubank.com, 2025-05-14). Nubank utilizes DCNv2 architecture for supervised fine-tuning of transaction embeddings, achieving improved AUC metrics through joint fusion of sequential and tabular data, surpassing traditional models like LightGBM

Paper Review: AlphaEvolve: A coding agent for scientific and algorithmic discovery (andlukyane.com, 2025-05-15). AlphaEvolve enhances coding agents' capabilities using iterative edits and evaluator feedback, optimizing tasks like data center scheduling and discovering novel algorithms, including a breakthrough in matrix multiplication efficiency

I trained neural nets on large cardinal axioms (karagila.org, 2025-05-14). Asaf Karagila details his experience training neural nets on large cardinal axioms, highlighting recent developments like the Bagaria–Goldberg characterization and summarizing a course on related concepts, including ultraexacting cardinals

Optimal Brain Surgeon (leimao.github.io, 2025-05-14). Derives the Optimal Brain Surgeon algorithm, discussing properties of positive definite matrices, Taylor expansion for neural networks, and the Woodbury matrix identity, relevant for pruning methods in modern neural networks

Introductions (blog.raymond.burkholder.net, 2025-05-17). This tutorial covers Discrete Variational Autoencoders and Deep Reinforcement Learning, focusing on VAE's categorical latent space and the Proximal Policy Optimization algorithm for sequential decision-making

Embeddings, Projections, and Inverses (johndcook.com, 2025-05-14). Discusses embeddings, projections, inverses, and the Moore-Penrose pseudoinverse. Covers the embedding of 3D vectors into quaternions and their retrieval process, and examines adjoints and left/right inverses

A Tiny Boltzmann Machine (eoinmurray.info, 2025-05-15). Explore Boltzmann Machines, generative AI models from the 1980s, and learn about the Tiny Restricted Boltzmann Machine designed for browser use and its mechanisms like Gibbs sampling and energy-based training

Is AI Security Work Best Done In Academia or Industry? Part 1 (cacm.acm.org, 2025-05-15). The debate between academia and industry in AI security research highlights the advantages of vast compute resources, data availability, and better compensation while emphasizing creativity's role in groundbreaking advancements

Workshop on Predictions and Uncertainty at COLT 25 (cstheory-events.org, 2025-05-14). The workshop at COLT 25 in Lyon, France invites researchers to explore advancements in uncertainty modeling and predictions in machine learning, covering topics such as Conformal Prediction and Risk-Averse Decision Making

DeepHyper: A Python Package for Massively Parallel Hyperparameter Optimization in Machine Learning (joss.theoj.org, 2025-05-19). DeepHyper is a Python package designed for massively parallel hyperparameter optimization in machine learning, supporting techniques such as multi-fidelity neural architecture search and utilizing high-performance computing resources

X X^t can be faster (arxiv.org, 2025-05-16). $XX^t$ offers a potential for improved speed in computational processes, leveraging advancements in data structures and algorithms, pivotal in various applications across computer science fields

On the Role of Weight Decay in Collaborative Filtering: A Popularity Perspective (arxiv:cs, 2025-05-16). Weight decay is crucial for collaborative filtering models, encoding popularity in embedding magnitudes. The proposed PRISM simplifies training, improving performance by 4.77% and reducing training time by 38.48% compared to existing strategies

Birch SGD: A Tree Graph Framework for Local and Asynchronous SGD Methods (arxiv:math, 2025-05-14). Birch SGD proposes a graph-based framework for distributed SGD methods, analyzing eight new algorithms with optimal computational complexity, revealing shared iteration rates, and different trade-offs for performance and communication efficiency

Scalable Computations for Generalized Mixed Effects Models with Crossed Random Effects Using Krylov Subspace Methods (arxiv:cs, 2025-05-14). Novel Krylov subspace methods improve computational efficiency for generalized mixed effects models with crossed random effects, achieving runtime reductions of up to 10,000 times compared to Cholesky decompositions, implemented in GPBoost

The Power of Random Features and the Limits of Distribution-Free Gradient Descent (arxiv:cs, 2025-05-15). The study reveals that distribution-free learning in neural networks using mini-batch stochastic gradient descent allows for polynomial-sized approximations with random features, introducing a new framework called average probabilistic dimension complexity

Diffusion Recommender Models and the Illusion of Progress: A Concerning Study of Reproducibility and a Conceptual Mismatch (arxiv:cs, 2025-05-14). A study on modern Denoising Diffusion Probabilistic Models reveals persistent methodological issues in top-n recommendation, indicating these complex models are generally outperformed by simpler counterparts, questioning their suitability for the task