Neuroscience-Based AI at Stanhope AI

Renowned cognitive and computational neuroscientist Rosalyn Moran founded Stanhope AI, which successfully raised £23 million in seed funding (approximately 40 billion KRW) earlier this year (Article 1, Article 2). Currently, she is a professor of computational neuroscience at King’s College London (KCL) and the director of the AI Institute at KCL. Having previously worked with Professor Karl Friston on the free energy principle and active inference, her research focuses on computational neuroscience, computational psychiatry, and neurology. She is now particularly focused on the intersection of deep neural networks, Bayesian inference (variational principles), and experimental neurobiology (neuroimaging and cognitive tasks).

Stanhope AI aims to develop AI systems inspired by the human brain. Their AI’s defining feature is that it “does not learn” in the conventional sense. Instead, as proposed by active inference, they design and fine-tune hierarchical generative models. This approach aims to reduce hallucination risks in AI while cutting costs. The company’s research revolves around the following key questions:

1. How can an autonomous machine find its own data?
2. How can an AI agent explain its beliefs and plans for action?

In pursuit of these answers, the company has developed an active inference model for visual search and created a Gym package that allows algorithms to interact with their environment through trial and error. Additionally, they are exploring how the brain’s locus coeruleus helps AI agents flexibly update and modify their world models in response to new experiences.

Based on this research, Stanhope AI has engaged in several projects. They are collaborating with the German government’s disruptive innovation agency, SPRIN-D, to explore problem-solving in previously unencountered environments. In collaboration with the British Navy, they have successfully demonstrated next-generation AI technology for drones, allowing them to explore unknown territories with minimal data.

Located in London, Stanhope AI comprises experts in neuroscience and robotics. Their goal is to apply theoretical neuroscience models to engineering, demonstrating strong generalization performance, energy efficiency, and robustness to quickly adapt to market demands. The company’s founding team includes Karl Friston, professor of theoretical neurobiology at University College London, and Biswa Sengupta, an engineer and theoretical computer scientist.

Stanhope AI is currently recruiting summer interns! (Link)

---

Developing human-level AI based on neuroscience principles has always been one of my dreams. Seeing efforts like Stanhope AI actively driving change in the industry and securing significant investment is both exciting and promising. However, it raises questions: What areas should I focus on moving forward? Are there aspects that they might be overlooking? These thoughts bring new inspiration and a lot to consider as I think about the direction of my own research.