Merve Noyan

🚀 Machine Learning Time Complexity: A Guide for MLOps Engineers (Optimize your models before they drain your compute budget) ✅ Decision Trees & Random Forests Efficient for many tasks, but deeper trees and large datasets can degrade performance. ✅ Neural Networks Highly powerful, but require heavy computation during training and inference. ✅ K-Nearest Neighbors (KNN) Training is fast, but inference slows as dataset size grows due to distance calculations. ✅ K-Means Clustering Great for unsupervised learning, but complexity increases with more clusters and iterations. ➡️ Why it matters for MLOps: Understanding time complexity helps reduce inference latency, optimize training pipelines, and manage compute resources effectively. Use this as your cheat sheet to deploy smarter, faster models. ⚡ 🎓 Learn Machine Learning & MLOps – Free Courses 🔗 Generative AI for Data Scientists → https://lnkd.in/dRYW2t26 🔗 IBM Data Science → https://lnkd.in/dhtTe9i9 🔗 SQL Basics for Data Science → https://lnkd.in/d6-JjKw7 🔗 Machine Learning Engineer – Career Academy → https://lnkd.in/dVGrGzBe 🔗 Top Python Machine Learning Libraries → https://lnkd.in/dTdWqpf5 📌 Save this visual 🔁 Repost to support other data professionals 🌐 More at: https://lnkd.in/dJw7mE-x ✍️ Image Credit: DataInterview.com #MachineLearning #MLOps #AI #DataScience #Python #FreeCourses #DeepLearning #ProgrammingValley #ModelOptimization #100DaysOfML #TechCareers

261

2 commentaires

BERT is back 😍 ModernVBERT SOTA-level embeddings (for its size) on visual-document retrieval (ViDoRe), matching models ~10× larger. WAIT BUT WHY? “Aren’t vision LLM based embeddings trained on more tokens?” Here’s the catch (retrieval ≠ generation): - Most recent VLM retrievers (e.g., ColPali) are built on causal decoders (autoregressive = next-token prediction; causal attention = left-to-right mask). - Great for generation, but suboptimal for retrieval, where we need globally contextual, comparable vectors. - Encoders use bidirectional attention (see all tokens at once) and train with contrastive objectives (pull matches together, push non-matches apart)—exactly what ANN similarity (vector DB) search scores. So what do you do? Train a dual encoder—twice. - Multi-vector (Late Interaction, ColBERT/ColPali-style): keep many token/patch vectors per item; score with MaxSim (each query token takes its best doc match) → fine-grained alignment on cluttered pages. - Single-vector (bi-encoder): pool to one embedding per item; score with cosine/dot-product → tiny indexes, fast CPU ANN; add a reranker for precision. result is ModernVBERT - A 250M-parameter vision–language encoder trained from the ground up for retrieval. Recipe - Text encoder: Ettin-150M (ModernBERT family; bidirectional) - Vision encoder: SigLIP2-B/16 (contrastive ViT; strong image–text alignment) - Modality alignment: MLM (masked language modeling; early fusion of patches into token space) - Finetuning: contrastive (InfoNCE, hard negatives) with both single-vector and Late Interaction heads Why it matters - Late Interaction loves bidirectionality → +10.6 nDCG@5 vs causal backbones on doc retrieval. - CPU-friendly query encoding at SOTA-level accuracy (for its size) → practical for RAG at scale. - Data-efficient: higher training resolution helps; text-only pairs transfer to visual-doc retrieval. When to use which 🧾 Scanned PDFs / tables / forms / screenshots / multi-column layouts → Late Interaction (e.g., ModernVBERT/colmodernvbert) for max recall on cluttered pages. ⛰️ Huge corpora / strict latency & memory / CPU-first serving → Single-vector bi-encoder (e.g., ModernVBERT/bimodernvbert or ModernVBERT/modernvbert-embed) + cross-encoder rerank for best cost/throughput. 🔀 Hybrid → single-vector for fast candidate retrieval, then re-rank top-K with Late Interaction (or a cross-encoder) when you need extra precision. Open models (MIT) on 🤗: ModernVBERT/colmodernvbert (Late Interaction) ModernVBERT/bimodernvbert (bi-encoder) ModernVBERT/modernvbert-embed Deep dive & benchmarks: https://lnkd.in/g9abnmHg

102

1 commentaire

💻 Why we still need small language models – even in the age of frontier AI Following a focused six-week sprint, our Project t0 team showed that small, open-weight language models can achieve near-frontier reasoning performance when applied to focused, domain-specific tasks.  By combining retrieval-augmented generation, reasoning trace fine-tuning and budget forcing at inference time, the team has built a 3B parameter model capable of tackling real-world health queries – and is small enough to run locally on a laptop. Federico Nanni, James Geddes, Ryan Chan and Tomas L. from the project team explore the details in this blog: https://bit.ly/4mfv4t7 To support others working on lean model deployment, we’re publishing: ✅ Two technical reports documenting our pipeline, data generation, fine-tuning on HPC infrastructure and evaluation process. ✅ Our code and trained models, open-sourced for reproducibility and reuse. ✅ The code for setting up a simple interface for multi-turn conversational interaction, allowing users to test the model in a realistic setting. This work is designed to support teams in compute-constrained or privacy-sensitive environments to build practical AI assistants that don’t require frontier-scale resources. 🔎 Learn more about our work and read the technical reports: https://bit.ly/4mfv4t7 #AI #ArtificialIntelligence

245

7 commentaires

Starting today, Claude Sonnet 4 supports 1 million tokens of context on the Anthropic API—5x more than before. This enables developers to process over 75,000 lines of code or dozens of research papers in a single request, ideal for data-intensive uses like long-running agents. Long context support for Sonnet 4 is now in public beta, starting with API users on Tier 4 and custom rate limits. Broader availability on our API will be rolling out over the coming weeks. It’s also available in Amazon Bedrock, and is coming soon to Google Cloud's Vertex AI. Learn more about Sonnet 4 and the 1M context window: https://lnkd.in/e-tfCYBg

6 640

434 commentaires

🤔 Wondering if the new Astro IDE already recognizes the latest Airflow 3.1 features? What about comparing Astro IDE's grasp of Airflow, versus other popular LLMs and coding agents like ChatGPT with gpt-5, Cursor with gpt-5, and Cursor with claude-4.5-sonnet? Tamara Janina Fingerlin dives into which of these solutions can help you write the best human-in-the-loop Apache Airflow DAG in her new blog post: https://bit.ly/3KLMf89

57

1 commentaire

Automate detection of unreliable LLM outputs by combining MLflow tracing with Cleanlab’s Trustworthy Language Models (TLM). 🚀 This blog post covers: ✅ Setting up MLflow to capture complete LLM interactions, including system prompts. ✅ Retrieving traces and efficiently batch-evaluating them with TLM to get trustworthiness scores and explanations—without ground truth labels. ✅ Tagging each MLflow trace with its trust score and explanation for ongoing review. ✅ Using MLflow Evaluation with a custom TLM metric to systematically compare model outputs. For the full guide and code samples, read more ➡️ https://lnkd.in/eHNWQrrK cc Chris Mauck, Data Scientist at Cleanlab #opensource #llm #oss #mlflow #cleanlab

59

1 commentaire

Does ML Zoomcamp actually change how you work? Alexander Daniel Rios, one of the course graduates, just published a reflection on what shifted for him: ➡️ Before: chaotic notes, unclear validation, jumping between models. ➡️ After: structured, end-to-end practice: 🔹 Systematic data prep and reproducible pipelines 🔹 Proper validation for regression/classification and metric selection tied to the use case 🔹 Confidence with NNs & transfer learning 🔹 Lightweight deployment (Flask, Docker; Lambda when needed) 🔹 Habit of sharing work in public (weekly LinkedIn/blog posts) What he built along the way: 🔹 Capstone 1: Blood-cell classifier + segmentationv(with a simple UI) 🔹 Capstone 2: Disaster-tweet classifier (spaCy, TensorFlow, BERT) 🔹 Kaggle: ML Zoomcamp 2024 Retail Forecasting with 4th place globally among dozens of participants The course provided him with a clear path from theory to pipeline to deployment. The community and public sharing turned weekly progress into feedback, connections, and even offers. Thank you, Alexander, for sharing your process! Read more about it here: https://lnkd.in/gSh93dv6

83

4 commentaires