Abstract
The human brain has immense learning capabilities at extreme energy efficiencies and scale that no artificial system has been able to match. For decades, reverse engineering the brain has been one of the top priorities of science and technology research. Despite numerous efforts, conventional electronics-based methods have failed to match the scalability, energy efficiency, and self-supervised learning capabilities of the human brain. On the other hand, very recent progress in the development of new generations of photonic and electronic memristive materials, device technologies, and 3D electronic-photonic integrated circuits (3D EPIC) promise to realize new brain-derived neuromorphic systems with comparable connectivity, density, energy-efficiency, and scalability. This paper argues the possibility of reverse-engineering the brain through architecting a prototype of a brain-derived neuromorphic computing system consisting of artificial electronic, ionic, photonic materials, devices, and circuits with dynamicity resembling the bio-plausible molecular, neuro/synaptic, neuro-circuit, and multi-structural hierarchical macro-circuits of the brain based on well-tested computational models. We further argue the importance of bio-plausible local learning algorithms applicable to the neuromorphic computing system that capture the flexible and adaptive unsupervised and self-supervised learning mechanisms central to human intelligence. Most importantly, we emphasize that the unique capabilities in brain-derived neuromorphic computing prototype systems will enable us to understand links between specific neuronal and network-level properties with system-level functioning and behavior.
Introduction
Despite the success of modern AI, current hardware remains fundamentally limited compared to the human brain. The brain operates on roughly 20 Watts of power while managing trillions of synaptic connections; in contrast, training a model like GPT-3 requires millions of dollars in energy and specialized GPU clusters. This paper proposes a paradigm shift from "brain-inspired" to "brain-derived" neuromorphic computing by integrating photonic, electronic, and ionic technologies into 3D integrated circuits (3D EPIC).
The authors identify four fundamental "Gaps" in current neuromorphic research:
- Principles: A lack of understanding of how intelligence emerges from networked, structured connectivity.
- Materials: Conventional silicon cannot faithfully replicate the ionic and molecular dynamics of biological dendrites and synapses.
- Devices: Standard CMOS transistors fail to emulate the complex temporal dynamics seen in biological neurons.
- Interconnects: Electronic wiring cannot scale to the ~8,000 synaptic connections per neuron found in the brain without massive energy penalties.
Methodology: The 3D EPIC Platform
The proposed 3D Electronic-Photonic Integrated Circuit (3D EPIC) platform is a multi-layered architecture designed to mimic the brain’s physical and functional hierarchy. This approach moves beyond simple inspiration to derive hardware properties directly from biological principles.
Materials & Devices
The system utilizes "intelligent" materials—including ferroelectrics, MXenes, and ionic-graphene—to create artificial synapses and dendrites. Electrochemical synapses use ion insertion, similar to neurotransmitters, to achieve non-volatile memory with high signal-to-noise ratios. Furthermore, the paper emphasizes that dendrites are active computing elements; by emulating dendritic branching with multi-gate FETs, the system can process spatiotemporal signals far more efficiently than a flat neural network.
Photonics for Connectivity
To solve the "wiring bottleneck," the system uses photonics for long-range communication. Through-Silicon Optical Vias (TSOVs) and Wavelength Division Multiplexing (WDM) allow thousands of signals to travel across the chip with minimal loss and latency. While electronics are ideal for high-density local micro-circuits, photonics are essential for the "macro-circuits" of the brain, enabling massive fan-out.
The XCAL Algorithm
Instead of traditional backpropagation, which is biologically implausible, the system employs Extended Contrastive Attractor Learning (XCAL). This is a local, error-driven learning rule that allows the hardware to self-supervise and adapt in real-time based on local activity differences.
Evaluation and Applications
The paper moves away from standard static datasets in favor of a 3D VR Navigation testbed. The goal is to evaluate the "artificial brain" on tasks requiring real-time adaptation, such as object recognition, selective attention, and map-based navigation. These benchmarks test compositionality, inference, and abstraction—generalizing concepts like "finding food" across different environments.
Key results indicate that nanoscale optoelectronic neurons can achieve energy efficiencies of <1 fJ per spike. HSPICE simulations of neuron spiking behavior showed a near-perfect match with experimental results from prototyped optoelectronic neurons. The 3D EPIC approach aims to support the high connectivity of the brain (~8,000 synapses per neuron) at a scale of billions of neurons.
Conclusion
While individual components like ionic synapses and photonic MZIs have been demonstrated, integrating them into a wafer-scale "artificial brain" remains a significant engineering challenge. Future work will focus on scaling to multi-stack 3D systems and structuring circuits to specifically emulate brain regions like the hippocampus and prefrontal cortex. Ultimately, this prototype serves as a tool for neuroscience, allowing researchers to test theories about the biological brain by manipulating its artificial counterpart.
Summary updated