Search

[2610.31146] Rethinking Fisher Vectors for Second-order Multi-scale Deep Feature Aggregation in Scalable Scene Retrieval

A codebook-free method encodes second-order information of multi-scale CNN features within the Fisher Kernel framework to improve scalable scene retrieval, achieving better accuracy on INRIA Holidays and UKBench with only a modest rise in vector dimension.

What Is Feature Engineering? Techniques & Examples

Feature engineering turns raw data into informative inputs for machine learning, improving model performance and training speed through iterative cleaning, encoding, scaling, and feature creation while avoiding excessive columns that hinder learning.

Real-time fraud detection with Feature Views | Databricks Blog

The document describes using Databricks Feature Views to build real-time fraud detection, ensuring training and serving use the same feature definitions to avoid data leakage and skew; it covers feature design, Unity Catalog governance, and a dual-path architecture for training and online inference.

Dharma Field Extractor: Any Document In, Structured Fields in JSON Out

Dharma Field Extractor provides a 2B-parameter model that outputs structured JSON fields from documents, either by open extraction or schema-directed requests, suitable for downstream systems; it generalizes across document types using OCR and a LoRA-trained fine-tuning regime.

KBlueLeaf (Shih-Ying Yeh)

The document outlines a set of AI image prompt tasks, including refining prompts from tags and text, generating Danbooru tags, creating sketches from photos and prompts, and producing anime-style dragon girl images, presented as a demo space for a HomeDiffusionModel.

Dharma-AI Launches Field Extractor Document Processing Model — GazeWire

Dharma-AI introduces Field Extractor, a 2B-parameter model that turns document images into structured outputs without fine-tuning, using Open and Schema-directed modes with a LoRA adapter. It achieves 0.834 field-level F1 on 546 documents and offers cost efficiency when self-hosted on a single L40S GPU versus APIs.

trmz/flux2-klein-alpha · Hugging Face

Toolchain for FLUX.2 Klein Alpha: foreground extraction to RGBA with auto-background and object removal, using 4B/9B extractors, AlphaVAE, and safetensors; includes download and setup steps and six-step inference.

Azure-TDSP-ProjectTemplate/Docs/Project/Charter.md at master · Azure/Azure-TDSP-ProjectTemplate · GitHub

Explains data sources, data movement to Azure, data sampling for modeling, and the tools used for feature engineering, modeling, and deployment.

opensearch-project/opensearch-neural-sparse-encoding-asym-v1 · Hugging Face

OpenSearch asym-v1 pairs a small query encoder that runs at search time with a larger document encoder used offline during indexing. The model reports 0.555 average NDCG@10 on BEIR-13, with lower retrieval FLOPS than the other compared models that encode queries.

Data description | User interest modeling competition

Translated title. Original: 数据说明|用户兴趣建模大赛

Dataset description for a user interest modeling challenge, detailing visual, face, and text cover features; includes formats and examples for visual numpy features, face attributes, and tokenized/desensitized cover words, with notes on data handling and preprocessing.