NanoEdge AI Studio Release Notes

NanoEdge AI Studio v5.2

Overview

NanoEdge AI Studio v5.2 introduces new neural network architectures, large dataset handling, and LLM-friendly library exports.

New Features

CNN and LSTM Models

  • Two new model architectures — CNN and LSTM — are now available for Classification and Extrapolation projects.

Large Dataset Support

  • Datasets larger than 250 MB are now supported. Oversized datasets are automatically reduced using a representative sampling strategy that preserves the diversity and distribution of the original data.

LLM-Ready Library Details

  • Explore your library’s algorithm and preprocessing details with any LLM — JSON files and a prompt example are included in the downloaded library folder.

Bug Fixes

  • Emulator — Fixed a coherence issue with quadratic expansion between the Python and C implementations.

  • Benchmark — Fixed a crash occurring when the loss value was NaN during benchmarking.

  • Benchmark — Fixed incorrect performance values displayed in cross-validation results for Regression projects.

  • Various other bug fixes and stability improvements.

NanoEdge AI Studio v5.1

Overview

NanoEdge AI Studio v5.1 delivers significant performance improvements to embedded ML solutions, expanded hardware support with dedicated acceleration capabilities. This release includes a minor API breaking change due to architectural improvements in the library structure.

Performance & Core Improvements

Embedded Code Execution Optimization

  • Substantial performance gains in embedded code execution

  • Knowledge base now fully integrated within the library, eliminating external dependencies

  • Streamlined runtime architecture for reduced overhead

Note

The knowledge integration necessitates API modifications. See the Breaking Changes section below.

New Hardware Support

STM32U3 Target

  • Full support for STM32U3 microcontroller with dedicated hardware accelerator (HSP); see dedicated documentation page.

  • Generated libraries automatically leverage hardware acceleration capabilities

  • Optimized inference performance on accelerated targets

Cortex-R52 Series

  • Support added for Cortex-R52 series processors

  • Extends NanoEdge AI Studio capabilities to real-time processing applications

New Features

Customizable Search Space

  • Users can now manually select specific models and preprocessing functions

  • Ability to constrain or expand search space based on application requirements

Benchmark Queue Management

  • New queueing system for benchmark operations

  • Schedule multiple benchmark runs for sequential execution

Breaking Changes

Library API Update

With the knowledge base now fully integrated within the library, the initialization API has been simplified for both classification and anomaly detection workflows.

Classification, Regression & Outlier API Changes

Initialization 

// Old API
neai_classification_init(knowledge)

// New API (v5.1)
neai_classification_init(void)

The knowledge parameter is no longer required. Knowledge is now embedded directly in the library at build time, eliminating the need to manage and pass knowledge data structures.

Anomaly Detection API Changes

Initialization 

// Old API
neai_anomalydetection_init()
neai_anomalydetection_knowledge(knowledge)  // Separate call for pretrained models

// New API (v5.1)
neai_anomalydetection_init(bool use_pretrained)

The separate neai_anomalydetection_knowledge() function has been removed. Instead, specify at initialization whether to use a pretrained model or perform on-device learning:

  • neai_anomalydetection_init(true) — Use embedded pretrained model (no learning phase required)

  • neai_anomalydetection_init(false) — Perform on-device learning using neai_anomalydetection_learn()