Neurodelphus
What if AI learned from biology the right way?
Not metaphors. Not loose inspiration. Empirically derived computational architecture from living neural circuits.
The Problem with How We Build AI
Modern generative AI systems — the large language models, image generators, and video synthesis engines reshaping the world — are extraordinarily powerful and extraordinarily expensive. Training and running them consumes megawatts of power. Yet the human brain accomplishes comparable feats on roughly 20 watts.
The standard response is to build better hardware. We argue the bottleneck is deeper: architecture. Current neural networks borrow loose metaphors from biology — neurons, layers, weights — while discarding the properties that actually make biological circuits so efficient: state-dependent nonlinear dynamics, history dependence, unsupervised local learning, spontaneous activity generation, and context-sensitive adaptation.
We propose to go back to the biological source material and derive AI architectural principles empirically, from living tissue.
What Are BioAdaptiveRFs?
A Biological Adaptive Response Function (BioARF) is our term for the actual input-output transformation performed by a volume of living neural tissue — characterized by the governing equation:
R(t) = G[ I(t) | H(t), Θ(t) ]
Where R(t) is the response pattern, I(t) is the input, H(t) is the history of prior inputs, and Θ(t) represents learned parameters that evolve over time. G is a nonlinear mapping function.
This is not a transfer function — which would imply the system is linear, memoryless, and time-invariant. Neural tissue is none of these things. It learns from single exposures, extracts signal from 70% noise, generates meaningful spontaneous activity without any input at all, and degrades gracefully when damaged. These are not quirks — they are the computational signatures of a radically different and more efficient architecture than anything currently running generative AI.
The BioARF framework is our method for formally capturing these properties and translating them into simulation-testable architectural primitives.
Deriving a BioARF from living neural tissue for instantiation in neuromorphic hardware.
The Research Program
Our laboratory has established the experimental platform: 3D neural tissue specimens — organoids, organotypic slices, neural aggregates — cultured on high-density multi-electrode arrays and systematically probed with thousands of distinct spatiotemporal stimulation patterns. Specimens are tested individually, as multi-specimen assemblies linked through our Virtual White Matter software platform, and as virtual embodied entities connected to real-time simulated environments.
This ongoing work generates the empirical BioARF database. The theoretical and simulation program funded by this initiative will:
-
Develop a formal mathematical framework mapping empirically-measured BioARFs onto computational graph representations amenable to simulation
-
Identify which BioARF properties confer energy-efficiency advantages and under what conditions
-
Instantiate candidate architectures in simulation and benchmark against transformers and recurrent networks on generative AI tasks under strict energy-budget constraints
-
Derive theoretical bounds on the expressivity advantage conferred by dynamical systems with BioARF-class nonlinearity
Why This Is Unconventional
The standard approach in neuromorphic computing is to design circuits that are inspired by biological principles derived from decades of neuroscience literature, then validate them computationally. We invert this: we measure what living tissue actually does under systematic provocation, then ask what architectural rules could produce that behavior — and whether those rules, instantiated in simulation, outperform conventional AI architectures on efficiency grounds.
This approach was first articulated in our 2017 paper "Connecting the Brain to Itself through an Emulation" (Frontiers in Neuroscience, PMC5492113), which proposed routing a brain's output through a real-time dynamical emulation of its own circuitry and feeding it back — a closed loop between an organ and a synthetic model of itself. The key insight from that work carries forward here: the right target for neuromorphic and AI architectural design is not static weight replication but dynamical regime matching.
Team
Mijail D. Serruya, MD, PhD
Director, Raphael Center for Neurorestoration, Farber Institute for Neuroscience, Thomas Jefferson University
Co-founder, Cyberkinetics | Co-founder, Neurodelphus LLC Principal designer, first BrainGate human BCI trial
Alessandro Napoli, PhD, Lead Engineer, Raphael Center for Neurorestoration, Neurodelphus LLC Signal processing, real-time neural interface systems
Learn more about the Neurorestoration program at Jefferson: neurorestoration.jefferson.edu
Email: Mijail.Serruya@jefferson.edu
