Architecture

Project Structure

With the upcoming releases of versions 3.0 and 2.7, Brain4J has undergone significant changes and improvements, aimed at enhancing customizability, scalability, and openness.
The project is now organized into two primary modules:

• brain4j-core: Includes all components related to machine learning, such as model architectures, training pipelines, optimization algorithms, and learning utilities.
• brain4j-math: Provides low-level operations, including GPU acceleration, advanced matrix and tensor manipulations, dataset management, and high-performance computational utilities.

Model Weights Management

Starting from version 2.7, all model parameters (including weights and biases) are now managed through Tensor objects rather than being encapsulated in higher-level abstractions like Synapse or Neuron.
This transition brings notable advantages:

• Reduces RAM usage
• Improves flexibility and modularity.
• Better support for GPU and parallel computing.
• Enhances scalability for large models.

Batch Processing

To accelerate convergence and reduce training time, Brain4J supports dataset batching.

• Pre-2.7 Behavior: During each epoch, every sample in a batch would initiate a lightweight virtual thread. The system would wait for all propagations to complete before proceeding to the next batch.
• Post-2.7 Behavior: With the introduction of tensorized operations, entire batches are now merged into a single tensor. The neural network processes the batch simultaneously, significantly improving computational efficiency and memory utilization.

Next Steps

Check out Optimizing Inference