Neurogenesis AI Scientist

Inspiration

Modern neuroscience experiments are slow, manual, and highly dependent on human intuition. While AI is widely used to analyze data, it rarely participates in the scientific process itself — forming hypotheses, designing experiments, interpreting results, and knowing when to stop.

I wanted to explore a harder question:
Can an AI behave like a scientist, not just a tool?

Neurogenesis AI Scientist was inspired by this gap — applying Gemini-powered reasoning to real EEG neuroscience experiments and allowing the AI to autonomously plan, execute, critique, and terminate experiments using scientific judgment.

What it does

Neurogenesis AI Scientist is an autonomous AI-for-science system that:

• Designs EEG P300 classification experiments using Gemini reasoning
• Runs real neuroscience experiments on 20 EEG subjects
• Evaluates results with deterministic metrics
• Critiques outcomes scientifically
• Decides whether to continue or stop experimentation autonomously

The system does not optimize blindly. It explicitly reasons about performance plateaus and halts exploration when further tuning is scientifically unjustified.

How we built it

The project is composed of two tightly integrated layers:

1. Deterministic Neuroscience Engine

   • EEG preprocessing using MNE-Python
   • P300 epoch extraction (0–600 ms, baseline corrected)
   • Within-subject evaluation using Logistic Regression
   • Fully reproducible pipeline with identical results across runs

2. Autonomous AI Scientist Layer

   • Gemini model generates experiment plans
   • Executes experiments via a controlled interface
   • Analyzes metrics and produces scientific critiques
   • Maintains persistent experiment memory
   • Implements autonomous stopping logic

Gemini is used strictly for scientific reasoning, not for data processing.

Challenges we ran into

• Designing a safe autonomy loop without letting the AI modify code
• Ensuring strict determinism for scientific validity
• Managing API quota limits during autonomous iteration
• Creating a demo that works without requiring judges to use their own API keys

Each challenge was solved with explicit constraints, frozen memory states, and reproducible execution.

Accomplishments that we're proud of

• Built a fully autonomous AI scientist, not just an ML pipeline
• Ran real EEG experiments on human neuroscience data
• Demonstrated scientific stopping behavior
• Achieved production-grade security (no exposed API keys)
• Delivered a clean, judge-safe demo under real-world constraints

What we learned

• Autonomy in AI requires constraints, not freedom
• Scientific progress is as much about stopping as improving
• Deterministic systems are essential when AI makes decisions
• Gemini excels at structured scientific reasoning beyond chat use cases

What's next for Neurogenesis AI Scientist

• Extend beyond Logistic Regression to deep EEG models
• Introduce cross-subject generalization experiments
• Add multi-objective scientific reasoning (accuracy vs robustness)
• Apply the framework to other scientific domains beyond neuroscience

This project represents a step toward AI systems that participate in science, not just assist it.

Built With

• api
• gemini
• mne-python
• python
• scikit-learn