Machine Learning Engineer: 1st July 2025

Normalizing Flows Are Capable Generative Models (machinelearning​.apple​.com). TarFlow, a Transformer-based Normalizing Flows model, achieves state-of-the-art results in likelihood estimation and image generation with advanced techniques for improved quality

Using generative AI to do multimodal information retrieval (amazon​.science). GENIUS model enhances multimodal information retrieval using generative AI, outperforming traditional methods in speed and accuracy with innovative techniques like semantic quantization

Researchers Uncover Hidden Ingredients Behind AI Creativity (quantamagazine​.org). Study reveals diffusion models' creativity stems from architectural imperfections in denoising, linking AI creativity with local patch processing techniques

Lessons Learned After 6.5 Years Of Machine Learning (towardsdatascience​.com). Lessons on deep work, trend management, and data analysis from 6.5 years in machine learning, covering personal insights and practical research experiences

Normalizing Flows Are Capable Generative Models (machinelearning​.apple​.com). TarFlow, a Transformer-based Normalizing Flows model, achieves state-of-the-art results in likelihood estimation and image generation with advanced techniques for improved quality

Polars Boosted My Algorithm's Speed by 25x (john​.soban​.ski). Polars library enhances algorithm speed by 25x, outperforming Pandas with parallel processing and Apache Arrow in Reduced Columb Energy classification

The Small Data Showdown ‘25: Is it Time to Ditch Spark Yet?? (mwc360​.github​.io). An exploration of Spark, DuckDB, Polars, and Daft performance for small data workloads in Microsoft Fabric; assessing if Spark remains viable

Thank you, Databend; Hello, LanceDB (xuanwo​.io). Xuanwo transitions from Databend to LanceDB, focusing on open-source multimodal databases, Rust optimizations, and personal milestones in tech and life

Pipelining AI/ML Training Workloads with CUDA Streams (towardsdatascience​.com). Explore CUDA streams for pipelining AI/ML training workloads, enhancing performance in PyTorch models through concurrent GPU operations and data augmentation techniques

Predicting Average IMDb Movie Ratings Using Text Embeddings of Movie Metadata (minimaxir​.com). Predicting IMDb movie ratings using text embeddings from metadata; explores neural networks, LLMs, and data wrangling with Polars

What Are Vector Embeddings? And Why They Matter in AI (sefiks​.com). Explore vector embeddings, their role in AI, applications in facial recognition, image search, LLMs, and concepts like PCA for dimensionality reduction

MUVERA: Making multi-vector retrieval as fast as single-vector search (research​.google). MUVERA enhances efficiency in multi-vector retrieval using fixed dimensional encodings, transforming complex queries into single-vector maximum inner product searches

Building a dynamic inventory optimisation system: A deep dive (engineering​.zalando​.com). Dynamic inventory optimization system at Zalando using AI, MLForecast, LightGBM, and AWS tools to enhance e-commerce supply chain efficiency

#511: From Notebooks to Production Data Science Systems (talkpython​.fm). Catherine Nelson discusses transitions from exploratory data science to production workflows, emphasizing code quality, software engineering practices, and MLOps platforms

Pareto ML Deployments (gojiberries​.io). Instance-level routing enhances ML model deployment, preventing regressions while balancing operational complexity and accuracy improvements on datasets like Adult income

Training 10,000 Anomaly Detection Models on One Billion Records with Explainable Predictions (databricks​.com). DAXS methodology utilizes scalable ECOD algorithm on Databricks for anomaly detection in manufacturing, improving predictive maintenance and operational efficiency with explainable insights

Another Another Place. My NOVA 2025 Demo. (benjamin​.computer). NOVA 2025 demo tribute to Anthony Gormley, using photogrammetry, machine learning, shaders, and techniques for creating 3D visuals on Linux

Patients want to know what information an AI model considers (ehudreiter​.com). Adarsa Sivaprasad's research reveals patients prioritize understanding AI model feature considerations over operational mechanics, highlighting a need for tailored explanations in healthcare AI

New Dataset Using Deep Learning to Predict Permafrost Thaw Damage in the Arctic with Elias Manos (arcticdata​.io). Elias Manos develops a deep learning model for predicting permafrost thaw damage in Arctic communities, addressing infrastructure risk and environmental changes

Unlocking the Secrets of Bulk Metallic Glass with Graph Computing (themultidisciplinarian​.com). Explore bulk metallic glass properties, design challenges, and the role of graph computing and machine learning in optimizing advanced materials

PCA analysis of Futures returns for fun and profit, part #1 (qoppac​.blogspot​.com). Explore PCA analysis of futures returns, uncover risk factors, trading strategies, and the relationship between risk and return in systematic trading

Can We Profit from Disagreements Between Machine Learning and Trend-Following Models? (quantpedia​.com). Exploration of using ML and trend-following models for equity returns, identifying profitable discrepancies for tactical allocation

Logistic regression - a simple briefing (blog​.engora​.com). Exploration of logistic regression for binary classification, odds ratios, sigmoid function, maximum likelihood estimation using Python's scikit-learn, and extensions to multiple classifications

Vectorize ROC Curve for Bayesian Models (juanitorduz​.github​.io). Vectorizing ROC curve computation for Bayesian models using Gaussian processes and the moons dataset with PyMC, ArviZ, and scikit-learn

Calculus Phobic’s Introduction to Differentiable Programming (andersource​.dev). Introduction to differentiable programming using JAX, focusing on optimization techniques, gradient descent, and automatic differentiation principles for solving computational problems

False Causes, Meet Attractor Dimension (methodsblog​.com). Yair Daon explores causation in public health using attractor dimensions, proposing BCAD for rejecting false causal links in ecological data analysis

Get ready for the PDE-Transformer (ge​.in​.tum​.de). Introducing PDE-Transformer: a novel architecture for physics simulations, combining UDiT and SWin features, outperforming existing methods in diverse PDE datasets

Random Vector Functional Link (RVFL) artificial neural network with 2 regularization parameters successfully used for forecasting/synthetic simulation in professional settings: Extensions (including Bayesian) (thierrymoudiki​.github​.io). RVFL artificial neural network utilizes Bayesian and Ridge2 techniques for effective forecasting and simulation in professional domains using Python and R implementations

(Don't) Forget About It: Toward Pareto-Improving GD (gojiberries​.io). Exploration of catastrophic forgetting in machine learning and methods to reduce regression during model training with a focus on loss function modifications

DiLoCoX: A Low-Communication Large-Scale Training Framework for Decentralized Cluster (arxiv:cs). DiLoCoX framework enables decentralized training of 100B+ parameter models over slow networks, utilizing Pipeline Parallelism and Adaptive Gradient Compression for efficiency

Interpretable Representation Learning for Additive Rule Ensembles (arxiv:cs). Enhanced interpretable additive ensembles leverage learnable sparse linear transformations for rule conditions, improving model complexity and performance in predictive tasks across various datasets

dreaMLearning: Data Compression Assisted Machine Learning (arxiv:cs). dreaMLearning framework utilizes Entropy-based Generalized Deduplication for efficient machine learning, accelerating training, reducing memory usage, and enhancing scalability for diverse applications

MPipeMoE: Memory Efficient MoE for Pre-trained Models with Adaptive Pipeline Parallelism (arxiv:cs). MPipeMoE enhances Mixture-of-Experts training with adaptive pipeline parallelism for efficiency, achieving 2.8x speedup and reducing memory footprint by 47%

Demystifying Distributed Training of Graph Neural Networks for Link Prediction (arxiv:cs). Distributed GNN training for link prediction faces performance degradation due to graph partitioning and negative sampling; SpLPG mitigates this with reduced communication costs

Feature-Wise Mixing for Mitigating Contextual Bias in Predictive Supervised Learning (arxiv:stat). Feature-wise mixing mitigates contextual bias in predictive ML, achieving 43.35% bias reduction and MSE improvements across classifiers without needing explicit bias identification

Towards Interpretable and Efficient Feature Selection in Trajectory Datasets: A Taxonomic Approach (arxiv:cs). Taxonomy-based feature selection improves interpretability and efficiency in trajectory datasets by organizing features into geometric and kinematic categories, enhancing predictive performance

SuperSONIC: Cloud-Native Infrastructure for ML Inferencing (arxiv:cs). SuperSONIC enables scalable ML inference using Kubernetes and GPUs, optimizing resource utilization for high energy physics and astrophysics research at major observatories

Efficient and Reuseable Cloud Configuration Search Using Discovery Spaces (arxiv:cs). Discovery Space abstraction improves cloud resource configuration search efficiency, enabling knowledge transfer and reducing optimization time by over 90% for diverse workloads

Hierarchical Modeling and Architecture Optimization: Review and Unified Framework (arxiv:stat). Review of hierarchical, conditional, and mixed-variable input modeling; introduces meta and partially-decreed variables; features design space graphs; implemented in Surrogate Modeling Toolbox for Bayesian optimization

CLoVE: Personalized Federated Learning through Clustering of Loss Vector Embeddings (arxiv:cs). CLoVE is a robust algorithm for Clustered Federated Learning, optimizing model accuracy through clustering client loss vector embeddings for personalized learning across diverse datasets

Data Uniformity Improves Training Efficiency and More, with a Convergence Framework Beyond the NTK Regime (arxiv:cs). Uniform data selection enhances training efficiency and performance in LLMs, proving smaller pairwise distances slow gradient descent and decrease approximation errors

Unimodal Strategies in Density-Based Clustering (arxiv:cs). Density-based clustering strategies improve parameter tuning for high-dimensional data using unimodal properties and Ternary Search algorithms across NLP, Audio, and Computer Vision tasks

The kernel of graph indices for vector search (arxiv:cs). Introduction of Support Vector Graph (SVG) for vector search in metric/non-metric spaces, leveraging kernel methods, enhancing graph indices like HNSW and DiskANN

I've got a big favor to ask - keeping Blaze running isn't expensive, but it does all add up, so I'm asking readers like you to help, if you can.
That's why I'm launching a Patreon page!. Nothing flashy, just a way for folks who find value in these newsletters to chip in a little each month. In return, you'll get:

• Real say in how Blaze evolves — vote on new topics, features, topic curation ideas
• First dibs on merch (details still cooking)
• That warm fuzzy feeling knowing you're supporting something that saves you time and keeps you plugged into great tech writing

If you are getting value from blaze, checking this out would mean the world. And if you can't contribute, no worries—the newsletters keep coming either way, and you can follow along on patreon for free.
Thanks for reading and being part of this nerdy corner of the internet. All the best - Alastair.