the basis for a neural network at the circuit level. Once again, the capabilities discussed in subsection three above will be examined here, but now at the next architectural level of the circuit in the full compute stack model, specifically (1) dendrite receptors, (2) neuron nucleus (soma) processing, (3) neuron memory, (4) sensor interfaces, (5) motor function interfaces, (6) axon connections, (7) synapse transmitters, and (8) synapse spiking signaling. Simplifications in the spike implementation are used for artificial spiking neural network implementation, and the way voltage is handed will vary since biological and electronic circuits are very different at the materials level and the component and device level. These details are not covered in this paper. a. artificial neuron considerations Synapse transmitters and dendrite receptors interconnections at the circuit level can be created using a two-dimensional (2‑D) crossbar. This may not be the most efficient, but it can create a robust interconnection fabric. Axon connections can be accomplished using nanowires, but only within the 2‑D organization made available in chip electronics. Since sensor and motor function interfaces are more specialized interfaces for transmission, reception, and signaling, there will have to be separate circuit designs for each one. Memory can be accomplished using chip electronic circuit modules made of memristors, phase-change memory, valance-change memory, or RRAM devices. Neuron nucleus processing can be accomplished with CPU cores made available to a handful of neural circuits, an arithmetic logic unit, or a microcontroller circuit. Once a neural circuit is available, it is possible to make networks of neural circuits available (Christensen 2022). b. biological neuron considerations Neuroscience research has a basic understanding of what a neuron is and how it functions. It explores the various facets of neurons in humans and animals from various points of view. Various animal testing ranging from flies, cockroaches, mice, and zebrafish, have been reviewed in the literature (Kim 2018, Huang 2017, Ryu 2018, Mitani 2018. For example, Mitani et al., by inserting microprobes into a mouse’s brain, gained insight into the spiking nature of neuron firing from specific condition tasks (Mitani 2018). The biological neuron receives information from the dendrites, processes the inputs in the cell nucleus (soma), transmits responses through the axons, and connects with other neurons through the synapses. 4. Microarchitecture Now that materials, components, devices, and circuits are available, they can be brought together to form a low-level architecture, which will be called a microarchitecture. Within a localized function, a microarchitecture is generated to execute a specific activity. A microarchitecture has greater complexity than a circuit but is not a complete architecture for a specific activity. It forms an essential building block as the next step in forming a complete neural capability. Microarchitectures form multilayer neural networks. Circuits are combined to form higher-level functional systems. In this process, multilayer neural networks are generated. The formation of neural sensing, cognition, motor function, and control microarchitectures are examples of functions that can be captured with this layer. For artificial neural network considerations, simple single-stage artificial neural networks are a focus of microarchitectures. Neurons function together to form a microarchitecture. For biological neural network considerations, biological microarchitectures would be the first step in cognitive development, where basic neuron connections are formed. 5. System architecture System architectures form the next higher level of organization of function, interfaces, and interdependent operation of various microarchitectures. A system architecture can be considered a system of systems, or a system of microarchitectures that operate interdependently. In neural systems, a system architecture is used for neural systems that perform a complete system function, like sensing, motor control, or a computing module category. The integration of neural systems like neural sensing, cognition, motor function, and control are examples of what are aggregated as a systems-of-systems architecture in this layer. For artificial neural network considerations, multilayer neural networks are a focus of system architecture, where more complicated architectures are formed that have more capability. For biological neural network considerations, rather than just single neurons connecting to other ones, with a system architecture, there are multiple layers of neurons connecting that are adapting in a manner that has a more specific focused function that can repeatably be called upon. 6. Algorithms Algorithms can utilize system architectures to generate a close coupling of “hardware” and “software.” They form approaches that are tied together and can then be made available as cyber-physical application capabilities that will be discussed in the next section. Algorithms tie one or more system architectures together with other resources, including the interdependence of multiple algorithms. a. sensing Algorithms are generated from architectures and circuits for each type of sense: (1) visual, (2) auditory, (3) taste, (4) olfactory, or (5) tactile detection of pressure and temperature. b. motor response Motor response aggregates tactile sensing, motor operation, proper response replication, and control computing. All these elements work together to provide a tuned functional motor response. c. learning Learning takes training data fed into neural networks and tunes the response within its neurons to produce the desired result more effectively. Learning is closely coupled with spiking neural networks. Once the neural connections have been established and properly conditioned, they can be activated in a tuned manner by future spiking events coming into the neural network. d. computing Computing requires processing, memory, and input and output connections to be correctly in place. Spiking neural networks process computational requests, especially in the brain. 7. Applications Applications combine major functional categories like cognition, JOHANSEN Human brain function and the creation model 2023 ICC 298
RkJQdWJsaXNoZXIy MTM4ODY=