pereirarodrigo/vijnana

Vijñāna (from Sanskrit, "consciousness") is a proof-of-concept reinforcement learning system that, in essence, learns to learn what to be curious about. This repository accompanies my paper “Temporal Meta-Curiosity via Recurrent Self-Reinforcing Intrinsic Meta-Learning”.

At its core is the Recurrent Intrinsic Meta-Learner (RIML), an architecture integrating curiosity-driven reinforcement learning with meta-learning and temporal encoding. The agent operates without any extrinsic rewards, relying solely on recursive novelty signals to structure its behavior.

🧠 Philosophical Inspiration

The name Vijñāna draws from Yogācāra and Madhyamaka traditions in Buddhist philosophy:

• Vijñaptimātra: All reality is mere consciousness.
• Pratītyasamutpāda: All phenomena arise dependently - a philosophical precursor to emergence.

This project embraces the idea that intelligence may not be imposed from the outside, but may emerge from within through recursive curiosity and temporal deliberation.

🛠️ Implementation

Vijñāna leverages:

• Random network distillation (RND) for intrinsic reward generation
• CNN-GRU encoding for spatiotemporal representation
• RL² meta-learning for cross-episode adaptation
• No external rewards, no hardcoded objectives

The agent builds its own internal landscape of meaning over time - sometimes acting, sometimes refusing to act.

This is not a goal-seeking system.
It is an agent learning how to want.

⚠️ Note

This is an early-stage proof-of-concept.
Expect imperfections, strange behaviours, and - at times - eerily emergent ones.