Neural Redis is a Redis loadable module that integrates feed-forward neural networks as a native data type. It targets developers, particularly for mobile and web applications, aiming to simplify machine learning by enabling in-database data collection and online training. This allows for rapid iteration and integration of predictive capabilities into applications.

How It Works

Neural Redis implements fully connected feed-forward neural networks using the RPROP (Resilient Backpropagation) learning algorithm. It supports both regression and classification tasks. A key design choice is to compress data collection and training phases into a single API, allowing networks to be used while training is ongoing. It features automatic data normalization and basic overtraining detection.

Quick Start & Requirements

• Install: Load the neuralredis.so module into Redis unstable (default branch from GitHub).

  redis-server --loadmodule /path/to/neuralredis.so
  

  Alternatively, add loadmodule /path/to/neuralredis.so to redis.conf.
• Prerequisites: Redis unstable.
• Setup Time: Minimal, assuming Redis is already running.
• Docs: README serves as the primary documentation and tutorial.

Highlighted Details

• Supports online training in separate threads, allowing network usage during training.
• Includes automatic overtraining detection via the AUTOSTOP option.
• Provides commands for creating, observing (training data), running, classifying, and inspecting networks.
• Demonstrates use cases with examples like addition prediction, Titanic survival prediction, and sentiment analysis.

Maintenance & Community

The project was initiated as a personal hackathon project by Salvatore Sanfilippo (antirez). While there's no explicit mention of ongoing maintenance or community channels, the core implementation is by a well-known figure in the Redis community.

Licensing & Compatibility

The README does not explicitly state a license. Given the author and context, it's likely to be permissive, but this requires verification. Compatibility with commercial or closed-source applications would depend on the final license.

Limitations & Caveats

This is explicitly labeled as "alpha code" and may contain bugs or crash the Redis server. It currently only supports sigmoid activation functions and mean squared error loss. AOF persistence rewrite is not implemented. The implementation is considered basic and could be significantly improved.