cazala/synaptic
Synaptic
Synaptic is a JavaScript neural network library used for building, training, and executing neural networks in Node.js and the browser. It provides a framework for constructing architecture-free neural network topologies, a backpropagation training engine for weight optimization, and a toolkit for implementing recurrent neural network frameworks.
The library enables the design of custom first or second order network architectures without predefined constraints. It supports a variety of specialized models, including Long Short-Term Memory networks, Hopfield networks, Liquid State Machines, and Multilayer Perceptrons.
The capability surface covers model serialization via JSON, sequence data processing, and network inference execution. It includes tools for architecture performance benchmarking and allows for asynchronous training execution using worker threads.
Features
- Neural Networks - Provides a complete library for building, training, and executing neural networks natively in Node.js and browsers.
- Activation Functions - Calculates a neuron output state based on input connections and a configurable squashing function.
- Backpropagation - Implements backpropagation algorithms to optimize neuron connection strengths based on output error.
- Custom Network Topology Design - Provides a framework for constructing arbitrary first or second order network topologies without predefined constraints.
- Browser-Based Inference - Enables executing model training and inference directly within the client browser.
- Inference Execution - Processes input data through a trained network to produce a computed output or final classification.
- Layer Connectivity Mapping - Establishes signal paths between layers using mapping strategies like all-to-all, one-to-one, or self-connections.
- Layered Architectures - Groups multiple neurons into a single unit to simplify the creation and management of architectures.
- Network Inference Execution - Passes an input vector through the network layers to calculate and return a final result.
- Architecture-Free Topologies - Provides a framework for constructing arbitrary first or second order network topologies without predefined constraints.
- Custom Architectures - Supports the design of non-standard network topologies and connection patterns for specific data problems.
- Neural Network Trainers - Adjusts internal weights using a trainer to teach a model how to solve specific tasks.
- Recurrent Neural Networks - Provides architectures for processing sequential data and time-series prediction using self-connecting neurons.
- Training Engines - Provides a core system for executing training loops and optimizing weights via error-based learning.
- Backpropagation Training - Adjusts neuron connections through backpropagation by comparing actual output against a target value.
- Neural Network Libraries - Offers a comprehensive JavaScript library for the development, training, and execution of neural networks.
- Activation Functions - Provides configurable mathematical functions to introduce non-linearity into neural network neuron outputs.
- Content-Addressable Neural Memory - Implements Hopfield networks for storing patterns and retrieving the most similar stored state based on input.
- Gated Connection Modulation - Enables controlling signal flow between network layers using a third layer as a masking mechanism.
- Asynchronous Training Utilities - Executes the training process in a background worker to prevent the user interface from freezing.
- Model State Formats - Implements a JSON-serializable format for exporting and importing neural network states and weights.
- Functional Model Exports - Converts trained model weights into standalone executable JavaScript functions or JSON files.
- Multilayer Perceptrons - Implements feed-forward networks with multiple fully connected layers for learning non-linear mappings.
- Model State Serialization - Saves and loads trained network weights and configurations using JSON for deployment and portability.
- Activation Processing - Processes all neurons in a layer sequentially and returns their combined outputs.
- Property Configurations - Sets the squashing function and bias for all neurons within a layer simultaneously.
- Hopfield Networks - Implements content-addressable networks that remember patterns and return the most similar stored pattern for new inputs.
- Liquid State Machines - Generates networks with randomly connected neurons that transform time-varying input into spatio-temporal activation patterns.
- Long Short-Term Memory Networks - Employs memory cells and gating mechanisms to learn long-term dependencies in time series data.
- Model State Serializers - Serializes network weights and architecture configurations into JSON for portability and reuse.
- Inter-Network Connections - Links the output of one network to the input of another using all-to-all or one-to-one patterns.
- Secure Network Connectivity Modules - Uses a specialized network to mask or modulate the connections between two other networks.
- AI & Machine Learning - Architecture-agnostic neural network library.
- Machine Learning and Vision - Architecture-free neural network library.
- Frameworks de Machine Learning - Architecture-free neural network library for browser and Node.js.
