sign be implemented? Electronics have specific functions, but technologically sophisticated device physics are harnessed. There is interdependence to a degree where external power, temperature control, and proper placement of parts within a larger design must be done. 3. Comparison between biological and artificial neural networks Table 3 below highlights the differences between biological and artificial neural network systems. (1) Power for biological neural networks is only required when a computation activates a neuron. This results in significant power savings. Electronics in chips typically apply power constantly to electronic devices. Neuromorphic computing architectures are trying to move away from that paradigm. (2) Biological neural networks use chemical energy when computing is required. Electronics require electrical energy, which cannot easily be stored in an artificial neuron and activated when necessary. (3) Biological neural networks use a complete three-dimensional (3D) architecture such that neurons can connect with other neurons anywhere in a three-dimensional space. This eases access to a more significant number of neurons. Electronics in chips typically only can connect in a two-dimensional (2D) planar manner. With the stacking of chips within a part package, there is the possibility for 2.5 D stacking access. Still, this is limited in comparison to biological systems. (4) Input and output connections are grown in biological neurons, with the axons having synaptic connections reaching out as far as necessary to access more neurons. Artificial neurons use switching fabrics, like a 2D crossbar, that connect any input to any output in a matrix fashion. This gives lots of flexibility but requires lots of hardware. (5) Memory and computing (central processing unit or CPU) resources are embedded in the biological neurons. The number of resources per neuron is tailored for its operation being distributed across elements. To mimic this behavior, artificial neural networks must embed memory modules and microcontrollers in each neuron or neuron cluster. The device physics differs between biological and artificial neurons. (6) Biological neural networks utilize molecular signaling mechanisms, including phosphorylation. Artificial neural networks utilize semiconductor properties where control inputs can modify signals. (7) Regarding parts architectures, biological neural networks are primarily composed of neurons, with various classification types used. Artificial neural networks use a variety of integrated circuit parts that are aggregated together. (8) In terms of learning, biological neural networks recruit other neurons for groups that can be recalled for duty to accomplish a learned task across the distributed network. Artificial neural networks must allocate hardware and software resources to accomplish the desired function. (9) Connecting sensors to computational resources is done slightly differently. For biological neural networks, the whole architecture consists of neurons, starting with neurons adapted to connect to the sensors, connected to transport neurons in the central nervous system. These are then connected to computational neurons in a neural network generated to interpret sensor data. For artificial neural networks, read-out electronics take sensor data and route it via an interconnect matrix to a computational neural network preprogrammed to process the specific sensor data. (10) For biological neural networks, brain computational network actuator commands are relayed via central nervous system neurons to the actuator neurons connected to the muscle. Response data is sent back to the neural network forming a closedloop system. For artificial neural networks, the neural network feeds interconnect resources that connect to actuators controlling motors. Sensors send feedback signals back to the neural network. IV. DISCUSSION A. Summary This paper explored the human brain function and architecture. It introduced a way of comparing biological and artificial systems. This study examined the neuron and neural network architectures seeking to observe their construction, operation, optimization, and adaptability. Regarding methodology, the study sought to look at the neuron and neural network systems from an engineering perspective and leverage systems engineering tools as part of the assessment. From ACT Neural Net Microarchitecture Neural Network Microarchitecture Synapse Dendrite Axon Learning Input Output Figure 22. Neural network microarchitecture. JOHANSEN Human brain function and the creation model 2023 ICC 307
RkJQdWJsaXNoZXIy MTM4ODY=