pulsatrix

ExAI-first C++ deep learning library — explainability as a first-class property of the computation graph, not a post-hoc wrapper. Every relevance-bearing layer ships a real, cited, conservation-tested Layer-wise Relevance Propagation (LRP) rule alongside its forward/backward math — never a placeholder or a post-hoc explainer bolted on afterward.

This page is the Doxygen API reference's landing page — generated from include/'s doc comments only. For guides, worked examples, and build instructions, see the full documentation site and the GitHub repository.

Modules

• Deep Learning Modules and Layers — Tensor/autograd core, layers, optimizers, losses, device backends
• Interpretability — post-hoc explainers (Model-Agnostic Explainers, Gradient-Based Explainers)
• Reinforcement Learning — environments, agents, and RL training algorithms
• Mechanistic Interpretability — activation probing, sparse autoencoders, circuit graphs, GFlowNets
• Neuro-Symbolic Reasoning — differentiable fuzzy logic, a Datalog engine, and a neural-predicate bridge tracing relevance from a symbolic derivation back into the network
• Data Loading, Transformation & Validation — dataset/loader/transform core and per-modality datasets
• Evolutionary Computation — genetic algorithms, neuroevolution, evolutionary HPO, PBT, E-GAN
• Hyperparameter Optimization — search space, Bayesian optimization, bandit-based early stopping
• Visualization — native C++ (Dear ImGui + ImPlot) charts and dashboards, opt-in via PULSATRIX_ENABLE_VIZ

Where to go next

• Getting Started — build the library and run your first example.
• Recipes — small, runnable programs demonstrating one tool at a time.