StarlitPupils/PrAlmodel

Neural Genesis (Pr-Al) is an open-source research codebase that implements a novel brain-inspired cognitive architecture. It aims to let task- and environment-driven differentiation emerge from a Pr-Al molecular gradient-like axis rather than hand-crafted modules. Intended users are AI researchers, computational neuroscientists and engineers exploring alternative architectures. Key features: - Four-layer hierarchical predictive-coding network (each layer predicts at different horizons). - Pr (short-term) vs Al (global) functional differentiation and conflict-driven IZ reflection loop. - Implements five neuro mechanisms: Hebbian plasticity, dopamine reward, curiosity, sleep consolidation, lateral inhibition. - Deliberately independent design (no Transformer/SSM/RNN/CNN); CUDA-accelerated kernels and training/eval scripts included. - Experiments: wikitext-103 LM (93M params) PPL 3.2, anomaly detection AUC=0.99. Requirements and quick-start (PyTorch, CUDA GPU) are provided.