Machine Learning Engineer

Making Real-time ML Models more robust in adversarial scenarios: Practical Tips and Monitoring Considerations (building.nubank.com.br, 2025-03-24). This guide offers practical tips to enhance the robustness of real-time ML models against adversarial attacks, highlighting the importance of understanding feature importance and the role of monitoring tools like SHAP and decision layer metrics

Embedding-Based Retrieval for Airbnb Search (medium.com/airbnb-engineering, 2025-03-19). Airbnb implemented an Embedding-Based Retrieval (EBR) system to enhance search relevance, employing contrastive learning, a two-tower neural network, and Approximate Nearest Neighbor (ANN) techniques like IVF for optimized performance

next OWABI webinar [27 March] (xianblog.wordpress.com, 2025-03-24). Join the upcoming OWABI webinar on March 27, featuring Meïli Baragatti discussing Approximate Bayesian Computation with Deep Learning and Conformal Prediction, introducing the ABCD-Conformal method that eliminates reliance on summary statistics

Understanding R1-Zero-Like Training: A Critical Perspective (github.com, 2025-03-22). Critical examination of R1-Zero-like training focusing on base models, reinforcement learning, and minimalist techniques using the Dr. GRPO algorithm for optimizing model performance on MATH level questions with specific frameworks

Three Hundred Years Later, a Tool from Isaac Newton Gets an Update (quantamagazine.org, 2025-03-24). Amir Ahmadi and colleagues enhance Newton's optimization method, enabling efficient application on complex functions using arbitrary derivatives, expanding its scope and speed for solving logistics, finance, and machine learning challenges

The Prospero Challenge (mattkeeter.com, 2025-03-24). The Prospero Challenge involves rendering a 1024x1024 image from 7866 math expressions using Python and Numpy, encouraging optimization with tools like LLVM and JIT compilation

Beyond the Scoreboard: Rethinking AI Benchmarks for True Innovation (eliza-ng.me, 2025-03-23). Explores the limitations of ML benchmarks, emphasizing Goodhart's Law, biases in language models, and the need for diverse evaluations that prioritize reasoning and problem-solving over mere metric optimization

The logit lens can be deceptive if not used properly (soniajoseph.ai, 2025-03-19). The logit lens, while a convenient tool for investigating neural network representations, can mislead researchers due to alignment issues with output space, emphasizing the importance of linear probes for accurate internal representation analysis

The Quickest Way to an Existential Crisis in Your Math Education (justinmath.com, 2025-03-23). Skipping computational practice leads to confusion when tackling abstract math concepts like eigenvalues and eigenvectors, ultimately resulting in an existential crisis in mathematical understanding

Formal Verification for Machine Learning Models Using Lean 4 (github.com, 2025-03-23). Framework leveraging Lean 4 for formal verification of machine learning models, focusing on essential properties like robustness, fairness, and interpretability, with interactive features via a web interface

Pen and Paper Exercises in Machine Learning (2022) (arxiv.org, 2025-03-21). Exploring conceptual foundations of machine learning through pen and paper exercises, supported by resources from the Simons Foundation and featuring tools like BibTeX, Connected Papers, and various citation management systems

Optimizing High-Dimensional Oblique Splits (arxiv:stat, 2025-03-18). This work optimizes high-dimensional s-sparse oblique splits for tree models, revealing a trade-off between statistical accuracy and computational cost, while integrating these splits into random forests for enhanced performance

Diffusion-augmented Graph Contrastive Learning for Collaborative Filter (arxiv:cs, 2025-03-20). Diffusion-augmented Contrastive Learning (DGCL) enhances collaborative filtering by integrating diffusion models with Graph Contrastive Learning, generating diversified contrastive views while preserving node-specific characteristics and semantic coherence

Interpretable Feature Interaction via Statistical Self-supervised Learning on Tabular Data (arxiv:stat, 2025-03-23). Spofe integrates kernel principal components with sparse polynomial functions for interpretable feature extraction from tabular data, applying a rigorous multi-objective knockoff selection procedure for effective feature selection and statistical reliability

A novel gradient-based method for decision trees optimizing arbitrary differential loss functions (arxiv:stat, 2025-03-22). A novel gradient-based method for decision trees optimizes arbitrary differentiable loss functions, utilizing first and second derivatives, and demonstrates application in classification, regression, and survival analysis, including censored data

Visual Insights (Part 2) (golem.ph.utexas.edu, 2025-03-20). John Baez discusses striking mathematical images from his Visual Insight blog, featuring contributions from artists like Refurio Anachro and Greg Egan, and uses XHTML and MathML for visual representation

Understanding Numpy's einsum (eli.thegreenplace.net, 2025-03-22). Learn to utilize numpy.einsum for efficient multi-dimensional array operations using Einstein notation, matrix multiplication, and ordering output dimensions with practical examples from machine learning applications

Unraveling spectral properties of kernel matrices – II (francisbach.com, 2025-03-24). Explores spectral properties of kernel matrices using translation-invariant kernels, examining eigenvalue decay through Fourier transforms, densities, and non-asymptotic bounds across different mathematical domains

On the relationship between sigmoid, softmax and tanh (blog.itdxer.com, 2025-03-20). Explores mathematical relationships between sigmoid, softmax, and tanh activation functions in neural networks, highlighting their transformational equivalence and potential performance implications in machine learning models

Targeting Neurodegeneration: Three Machine Learning Methods for G9a Inhibitors Discovery Using PubChem and Scikit-learn (arxiv:q-bio, 2025-03-20). Three machine learning models developed using scikit-learn analyze G9a inhibitors, predicting efficacy, inhibition probability, and ranking functional groups with various metrics including accuracy of 78.1% and up to 17.81% mean relative error

Network Embedding Exploration Tool (NEExT) (arxiv:cs, 2025-03-20). NEExT is a tool for embedding collections of graphs using user-defined node features, offering interpretability and fast embedding through the Vectorizers library, suitable for both supervised and unsupervised graph analysis

Food Delivery Time Prediction in Indian Cities Using Machine Learning Models (arxiv:cs, 2025-03-19). This research predicts food delivery times in Indian cities using machine learning models like LightGBM, incorporating real-time factors such as traffic, weather, and local events, achieving an R2 score of 0.76

C(NN)FD -- Deep Learning Modelling of Multi-Stage Axial Compressors Aerodynamics (arxiv:cs, 2025-03-18). A generalized deep learning framework for multi-stage axial compressors predicts flow fields and performance using physics-based dimensionality reduction and a multi-dimensional physical loss function, offering interpretable and accurate results across various operating conditions

Fake Runs, Real Fixes -- Analyzing xPU Performance Through Simulation (arxiv:cs, 2025-03-18). xPU-Shark is introduced as a fine-grained performance analysis tool for ML models at the machine-code level, optimizing communication by 15% and reducing token generation latency by 4.1% through hardware-level simulation