A Computational Perspective on NeuroAI and Synthetic Biological Intelligence

TL;DR

The paper surveys NeuroAI with a central focus on SBI, outlining three interacting domains—hardware, software, and wetware—and presenting computational frameworks for integrating living neural tissue with digital systems. It details foundational biology (BNNs), hardware advances (neuromorphic computing), software models (neuro-symbolic AI, RL, and active inference), and SBI architectures (open/closed-loop, assembloids/connectoids), along with digital twins and data pipelines. It also discusses ethical, regulatory, and societal implications and highlights recent advances in disease modeling and biohybrid platforms that point toward energy-efficient, interpretable, and adaptive intelligent systems. The review argues that true progress lies in the seamless fusion of biological and artificial substrates, supported by standardized data practices and anticipatory governance to unlock translational impact in medicine and beyond.

Abstract

NeuroAI is an emerging field at the intersection of neuroscience and artificial intelligence, where insights from brain function guide the design of intelligent systems. A central area within this field is synthetic biological intelligence (SBI), which combines the adaptive learning properties of biological neural networks with engineered hardware and software. SBI systems provide a platform for modeling neural computation, developing biohybrid architectures, and enabling new forms of embodied intelligence. In this review, we organize the NeuroAI landscape into three interacting domains: hardware, software, and wetware. We outline computational frameworks that integrate biological and non-biological systems and highlight recent advances in organoid intelligence, neuromorphic computing, and neuro-symbolic learning. These developments collectively point toward a new class of systems that compute through interactions between living neural tissue and digital algorithms.

Figures (8)

• Figure 1: Representation of dendrite branches as electrical circuits with active elements.
• Figure 2: Graph representation of the relations between different fields that build the foundation for NeuroAI.
• Figure 3: Illustration of the neuromorphic architecture. A: Traditional von Neumann architecture with separate memory and processing units connected by a shared communication BUS, which creates a bandwidth bottleneck and increases energy consumption. B: Neuromorphic architecture composed of analog, digital, or hybrid cores activated by spike-based signals, mimicking the event-driven dynamics of biological neurons. C: Analog neuromorphic core based on memristors-resistive elements that simultaneously store and modulate synaptic weights without requiring a clock signal. D: Digital neuromorphic core with integrated local memory and processing units. These cores can operate with pre-trained weights or perform local asynchronous updates via embedded learning engines.
• Figure 5: Diagram of reinforcement learning. Green arrows indicate a model-based approach. Experience is stored in an offline buffer to enable off-policy learning or used directly for on-policy updates.
• Figure 6: Diagram of active inference over two timesteps. The agent updates internal beliefs about hidden causes and selects actions that minimize both current and expected free energy.