Skip to content

Latest commit

 

History

History
85 lines (55 loc) · 3.59 KB

changelog.md

File metadata and controls

85 lines (55 loc) · 3.59 KB

Changelog Explanation

This document provides insights into the reasoning behind the changes made between versions of the micro-library for creating minimal neural networks in JavaScript.

Version 7: JSON as Default Export with Optional Binary Compression

Changes:

  • Default Export Mode: JSON is now the default export format.
  • Binary Compression: Added optional binary compression for larger neural networks.

Reasoning:

  • JSON as Default: JSON is human-readable and widely supported, making it easier for developers to inspect, debug, and share models. This change prioritizes usability and interoperability.
  • Binary Compression: For larger models, JSON can result in large file sizes. Binary compression is introduced as an optional feature to reduce file size and improve load/save performance for larger neural networks. Compression is now reliable.

Version 6: Dramatic File Size Optimization

Changes:

  • File Size Reduction: A model with 18k parameters was reduced from 376kb to 73kb.

Reasoning:

  • Efficiency: Smaller file sizes improve load times and reduce memory usage, which is critical for web-based applications and environments with limited resources.
  • Optimization: We are now using .uai as default exporting method for binary compression.

Version 5: Metadata Updates

Changes:

  • New Metadata Fields: Added new metadata fields to provide more information about the model.
  • Removed "Layers" Metadata: The unnecessary "layers" metadata field was removed.

Reasoning:

  • Enhanced Metadata: Additional metadata fields improve the ability to document and understand the model's structure and training process.

Version 4: Optimized Save/Load Handling for .uai Extension

Changes:

  • Optimized Save/Load: Improved handling for saving and loading models with the .uai extension.
  • JSON-Based Extension: The .uai extension is now JSON-based.

Reasoning:

  • Performance: Optimized save/load operations enhance the library's usability, especially for larger models or frequent model updates.
  • Standardization: Using a JSON-based .uai extension ensures compatibility and readability, aligning with modern development practices.

Version 3: Code Rewrite and Adam Optimizer

Changes:

  • Adam Optimizer: Introduced the Adam optimizer for training.
  • Code Rewrite: Complete code rewrite for maintenance purposes.

Reasoning:

  • Adam Optimizer: Adam is a popular and effective optimization algorithm, offering better convergence and performance compared to simpler optimizers like SGD.
  • Code Rewrite: A complete rewrite improves code maintainability, readability, and extensibility, ensuring the library can evolve more easily in the future.

Version 2: Softmax, Cross-Entropy, and Classification Labels

Changes:

  • Softmax Activation: Added softmax activation for output layers.
  • Cross-Entropy Loss: Introduced cross-entropy loss for classification tasks.
  • Labels: Added support for classification labels.

Reasoning:

  • Classification Support: These changes enable the library to handle classification tasks, expanding its use cases beyond regression.
  • Standard Techniques: Softmax and cross-entropy are standard tools for classification, improving the library's alignment with machine learning best practices.

Version 1: Initial Release

Changes:

  • First Release: The initial version of the library.

This changelog reflects the library's evolution, focusing on usability, performance, and feature expansion while maintaining minimalism and simplicity.