Machine Learning Engineer: 10th June 2025

Deep learning gets the glory, deep fact checking gets ignored (rachel.fast.ai, 2025-06-03). Deep learning models, like Transformers, were used for enzyme function prediction, revealing significant errors and emphasizing the need for domain expertise and accurate data validation in biological AI applications

Beyond data poisoning in federated learning (gwolf.org, 2025-06-04). This research explores hyperdimensional data poisoning attacks in federated learning, employing cosine similarity to confuse neural networks, highlighting risks in machine learning from adversarial noise in training data

AI interpretability is further along than I thought (seangoedecke.com, 2025-06-07). Recent advancements in AI interpretability reveal promising tools like sparse autoencoders and replacement models, enabling insights into model decision-making and feature identification despite complexities like polysemanticity and internal superposition

Grokking, Generalization Collapse, and the Dynamics of Training Deep Neural Networks with Charles Martin - #734 (twimlai.com, 2025-06-05). Discussion with Charles Martin on Weight Watcher, an open-source tool based on Heavy-Tailed Self-Regularization theory, addressing deep neural network training phases like grokking and generalization collapse, and applications in generative AI

A Neat Not-Randomized Algorithm: Polar Express (ethanepperly.com, 2025-06-07). The Polar Express algorithm optimizes matrix sign methods using compositions of polynomials for efficient singular value transformations, crucial for GPU-based neural network applications, especially in scenarios avoiding expensive SVD computations

Machine learning with hard constraints: Neural Differential-Algebraic Equations (DAEs) as a general formalism (juliabloggers.com, 2025-06-03). Neural DAEs empower machine learning with hard constraints by integrating algebraic conditions directly into models, ensuring solutions respect required properties across simulations, utilizing DifferentialEquations.jl for effective problem-solving

scikit-learn, glmnet, xgboost, lightgbm, pytorch, keras, nnetsauce in probabilistic Machine Learning (for longitudinal data) Reserving (work in progress) (thierrymoudiki.github.io, 2025-06-06). Explore probabilistic Machine Learning models for insurance reserving using scikit-learn, glmnet, xgboost, lightgbm, PyTorch, Keras, and nnetsauce, focusing on predictions and IBNR estimates

Anomaly Detection (blog.raymond.burkholder.net, 2025-06-04). A novel cluster-aware causal mixer for online anomaly detection improves multivariate time series analysis by grouping channels based on correlations, utilizing dedicated embedding layers, and preventing false positives through accumulated anomaly evidence

reproducing deep double descent (stpn.bearblog.dev, 2025-06-04). Stephen Wan details his efforts reproducing deep double descent experimentation using ResNet18 on CIFAR-10, focusing on model size, noise levels, and the practical challenges faced during implementation

Model stacking doesn’t work (snimu.github.io, 2025-06-07). Extensive experiments revealed that model stacking, aimed at decentralized pre-training and utilizing techniques like tied embeddings, leads to worse performance than individual models, highlighting critical insights about norms and layer usage

Compressive algorithmic randomness: Gibbs-randomness proposition for massively energy efficient deep learning (science-memo.blogspot.com, 2025-06-03). Explores compressive algorithmic randomness and Gibbs randomness in deep learning, utilizing dual tomographic compression and inverse compressed sensing for significant energy efficiency improvements in model training

Fractal dimensions and self similar behaviour in neural networks (danmackinlay.name, 2025-06-03). Fractal behaviors observed in neural networks include fractal loss landscapes, SGD trajectory modeling using Feller processes, and connections to singular learning theory, employing techniques like topological data analysis and Hausdorff dimension metrics

Non-Heuristic Selection via Hybrid Regularized and Machine Learning Models for Insurance (arxiv:stat, 2025-06-05). Machine learning models, including CatBoost and regularization techniques like Lasso, effectively predicted travel insurance purchases, achieving an AUC of 0.861 and an F1 score of 0.808 using feature selection and k-fold cross-validation

carps: A Framework for Comparing N Hyperparameter Optimizers on M Benchmarks (arxiv:cs, 2025-06-06). carps is a benchmark framework for evaluating hyperparameter optimizers, supporting blackbox, multi-fidelity, and multi-objective tasks, with a library of 3,336 tasks and 28 optimizer families for comprehensive performance studies

Machine Learning Predictions for Traffic Equilibria in Road Renovation Scheduling (arxiv:cs, 2025-06-06). Machine learning-based surrogate models predict traffic congestion from simultaneous road renovations, outperforming traditional simulations. XGBoost achieves 11% MAPE, significantly reducing computational burdens in long-term maintenance planning

Learning Binarized Representations with Pseudo-positive Sample Enhancement for Efficient Graph Collaborative Filtering (arxiv:cs, 2025-06-03). BiGeaR++ enhances graph representation binarization for collaborative filtering by mitigating information loss, utilizing pseudo-positive samples, and achieving state-of-the-art performance improvements of 1%-10% across five datasets compared to BiGeaR