NeoML – Machine learning framework for deep learning and traditional algorithms (github.com)

ABBYY has released NeoML, a versatile end-to-end machine learning framework designed to support both deep learning and traditional ML algorithms. NeoML enables developers to build, train, and deploy models across a wide range of AI tasks including computer vision, natural language processing, OCR, and document layout analysis. Its support for over 100 neural network layer types alongside 20+ traditional algorithms like classification, regression, and clustering makes it a comprehensive tool for diverse ML workflows. Technically, NeoML stands out with robust cross-platform compatibility, running efficiently on Windows, Linux, macOS, iOS, and Android, with CPU and GPU acceleration where available. It supports multiple programming languages including Python, C++, Java, and Objective-C and offers ONNX model import functionality, facilitating interoperability with other ML frameworks. NeoML’s modular architecture separates the user interface from low-level math engines, allowing users to select optimized CPU or GPU engines based on hardware, with JIT compilation via xbyak enhancing performance on x86_64 processors. Although GPU training is supported on Windows, iOS, and Android, GPU acceleration is currently unavailable on Linux and macOS. Its comprehensive documentation and sample tutorials provide an accessible entry point for developers. This framework’s emphasis on thread-safe math engines, multi-language support, and device-specific acceleration addresses key practical needs for scalable, cross-device ML development. By combining traditional ML algorithms and advanced DNN features within a unified, performant platform, NeoML is a significant addition to the AI/ML toolbox, especially for applications requiring extensive document processing and multi-platform deployment.