Machine Learning Engineer: 3rd June 2025

Meet PICT: the differentiable Fluid Solver for AI & machine learning in PyTorch (ge.in.tum.de, 2025-05-28). PICT is a differentiable fluid solver for AI in PyTorch, enabling learning tasks like turbulence modeling and sub-grid scale model training while providing fast simulations with accurate gradients

A Story of A Million Rows: Building a Lightweight Batch Pipeline with Temporal and DuckDB (medium.com/riskified-technology, 2025-05-28). A lightweight batch-processing pipeline was crafted using Temporal and DuckDB to handle independent operations for millions of rows, streamlining tasks with gRPC and optimizing for scalability through partitioning

Fast Kernels (crfm.stanford.edu, 2025-05-28). Fast AI-generated CUDA-C kernels outperform expert-optimized PyTorch kernels. Utilizing KernelBench, the team demonstrates advanced techniques yielding significant performance improvements, suggesting a promising future for automated kernel generation

Why DeepSeek is cheap at scale but expensive to run locally (seangoedecke.com, 2025-06-01). DeepSeek-V3's performance tradeoff centers on throughput and latency, showing that high-batch inference optimizes GPU efficiency, but local execution is costly due to matrix multiplication inefficiencies and queuing delays

Will my apartment in 5th avenue be overpriced or not? Harnessing the power of www.techtonique.net (+ xgboost, lightgbm, catboost) to find out (thierrymoudiki.github.io, 2025-05-28). Explore pricing predictions for real estate using www.techtonique.net, leveraging advanced machine learning algorithms like XGBoost, LightGBM, and CatBoost for conformal regression analysis

Deep-learning based multi-modal models for brain age, cognition and amyloid pathology prediction (alzres.biomedcentral.com, 2025-05-31). A hybrid deep-learning framework utilizing 3D convolutional neural networks and demographic data predicts brain age, cognitive function, and amyloid pathology from over 10,000 MRI scans, enhancing early diagnosis of neurodegenerative diseases

Leaps in Thought (blog.raymond.burkholder.net, 2025-05-31). Exploration of diffusion models through associative memory reveals memorization-generalization dynamics and the emergence of spurious states akin to Hopfield networks, providing novel insights and empirical validation

Hands-On Attention Mechanism for Time Series Classification, with Python (towardsdatascience.com, 2025-05-30). Explore the application of attention mechanisms in time series classification using Python, focusing on detecting anomalies in sine waves through dynamic attention scores and a bidirectional LSTM model

Learning Recommendation Systems: Bilateral Variational Autoencoder (BiVAE) (fanyangmeng.blog, 2025-05-29). BiVAE is a novel recommendation algorithm that uses bilateral treatment of user-item interactions, enhancing collaborative filtering through symmetrical learning and addressing limitations of traditional VAE approaches

Announcing Think Linear Algebra (allendowney.com, 2025-05-28). Allen Downey announces 'Think Linear Algebra', a book utilizing Jupyter notebooks, case-based learning, and tools like NumPy and SciPy while emphasizing the conceptual tools of linear algebra in problem-solving

Learning Recommendation Systems: Alternating Least Squares (ALS) (fanyangmeng.blog, 2025-05-27). Master the ALS algorithm with a comprehensive guide from mathematical foundations to PySpark implementation. Build production-ready recommendation systems with code examples and parameter tuning using the MovieLens dataset

Using ‘Slop Forensics’ to Determine Model Ancestry (dbreunig.com, 2025-05-30). Drew Breunig explores slop forensics, a tool by Sam Paech, to analyze model ancestry through slop profiles, revealing shifts in data generation approaches among large language models

Maximum Likelihood estimation with Quipu, part 1 (mathias-brandewinder.github.io, 2025-05-28). Learn how to use Quipu, an F# implementation of the Nelder-Mead algorithm, for Maximum Likelihood Estimation to analyze equipment failure data, leveraging libraries like MathNet.Numerics and Plotly.NET

Uncertainty Estimation with Conformal Prediction (m-clark.github.io, 2025-06-01). Explore uncertainty estimation techniques, focusing on conformal prediction, a model-agnostic method for generating prediction intervals, and comparing it with traditional approaches while highlighting the importance of understanding prediction confidence

Permutations and SHAPley values for feature importance in techtonique dot net’s API (with R + Python + the command line) (thierrymoudiki.github.io, 2025-06-01). Explore feature importance in machine learning using permutations and SHAP values with techtonique.net’s API via R, Python, and command line, covering model agnosticism and interpretability techniques

Bayesian Optimization for Hyperparameter Tuning of Deep Learning Models (towardsdatascience.com, 2025-05-27). Explore how Bayesian Optimization surpasses Grid Search in efficiency and performance for hyperparameter tuning in deep learning models, utilizing Keras and advanced techniques like surrogate models and acquisition functions

Tannaka Reconstruction and the Monoid of Matrices (golem.ph.utexas.edu, 2025-06-01). John Baez and Todd Trimble explore Tannaka reconstruction applied to the monoid of n x n matrices, demonstrating that finite-dimensional algebraic representations form a free 2-rig with subdimension properties

Information Processing Complexity as Spacetime Curvature: A Formal Derivation and Physical Unification (novaspivack.com, 2025-06-01). A rigorous proof connecting information processing complexity to local spacetime curvature via Landauer's principle and Fisher information geometry, introducing the 'Information Complexity Tensor' as a key component in Einstein's field equations

New Mathematical Framework Better Illustrates Complex Data Patterns (datascience.ucsd.edu, 2025-05-27). Researchers at UC San Diego develop a new hierarchical clustering framework utilizing Lebesgue integration, facilitating meaningful data grouping across varying population densities, and addressing modern data science analytical needs

Learning Schrödinger bridges (danmackinlay.name, 2025-05-29). This notebook explores Schrödinger bridges as stochastic bridge processes to condition neural denoising diffusion models, highlighting connections to optimal transport and referencing multiple relevant works in generative modeling and score matching

Comparative Analysis of Machine Learning Models for Lung Cancer Mutation Detection and Staging Using 3D CT Scans (arxiv:cs, 2025-05-28). Comparative performance of FMCIB+XGBoost and Dinov2+ABMIL for KRAS and EGFR mutation detection and cancer staging in 3D lung CT scans shows supervised models excel in mutations while SSL aids staging generalization

Speeding up Model Loading with fastsafetensors (arxiv:cs, 2025-05-29). Fastsafetensors optimizes tensor deserialization in safetensors files by directly instantiating on-disk parameters in device memory, achieving performance improvements of 4.8x to 7.5x for loading large models like Llama and Bloom

PolarGrad: A Class of Matrix-Gradient Optimizers from a Unifying Preconditioning Perspective (arxiv:stat, 2025-05-27). Introducing PolarGrad, a matrix-gradient optimizer leveraging polar decomposition, improving convergence over Adam and Muon in deep learning tasks by addressing curvature and gradient anisotropy through structure-aware preconditioning methods

Taming Hyperparameter Sensitivity in Data Attribution: Practical Selection Without Costly Retraining (arxiv:stat, 2025-05-30). This study explores hyperparameter sensitivity in data attribution methods, emphasizing the challenge of costly model retraining and proposing efficient tuning strategies and a lightweight regularization selection procedure for influence function methods

Deep k-grouping: An Unsupervised Learning Framework for Combinatorial Optimization on Graphs and Hypergraphs (arxiv:cs, 2025-05-27). Deep k-grouping introduces an unsupervised learning framework utilizing one-hot encoded polynomial unconstrained binary optimization and GPU-accelerated algorithms for efficient solving of large-scale k-grouping problems in graphs and hypergraphs

GARLIC: GAussian Representation LearnIng for spaCe partitioning (arxiv:cs, 2025-05-30). GARLIC utilizes Gaussian representation learning for effective high-dimensional search and classification, achieving fast building times and superior recall rates, outperforming existing methods like Faiss-IVF in k-NN retrieval and classification tasks

Data Model Design for Explainable Machine Learning-based Electricity Applications (arxiv:cs, 2025-05-29). A taxonomy is proposed to structure multivariate data in energy applications, enhancing machine learning model development. It evaluates features' impact on forecasting accuracy using interpretable techniques and feature importance methods