Back-Gate Switching Circuit for Low-Power Product-Sum Operation

The patent divides the artificial neural network into multiple layers (input layer, hidden layers, output layer) with neurons organized in hierarchical structures. Each neuron circuit is segmented into distinct functional components (synaptic weight circuits, activation function circuits, summation circuits) that can be independently optimized. This segmentation allows the system to achieve high computational capability through modular organization while managing power consumption at the component level.

The patent implements periodic action through the sequential processing architecture where neurons in each layer are activated in discrete time steps or cycles. The forward propagation process operates periodically, with input signals being processed layer by layer in a rhythmic manner. This periodic operation allows for efficient resource utilization and reduces instantaneous power demands compared to continuous parallel processing of all neurons simultaneously.

By segmenting the neural network into spatially distributed layers and neurons, the patent distributes heat generation across multiple localized regions rather than concentrating it in a single high-density processing unit. Each neuron circuit generates heat locally, but the distributed architecture allows for better thermal management through spatial separation, reducing the overall temperature impact on any single circuit component.

The patent introduces intermediary elements such as activation function circuits and summation circuits that act as buffers between synaptic weight circuits and output neurons. These intermediary components distribute and regulate signal flow, thereby distributing and regulating heat generation patterns. The intermediary structures allow thermal energy to be dispersed across additional circuit elements, reducing peak temperatures and minimizing the impact on critical circuit characteristics.

The patent segments the large-scale neural network into manageable layers, with each layer containing a specific number of neurons and connection strengths. This hierarchical segmentation transforms an overwhelming monolithic structure into organized, modular units that can be designed, analyzed, and optimized independently. The segmentation principle allows the system to handle large numbers of circuits by organizing them into structured layers, making the overall complexity tractable while maintaining high processing power.

The patent addresses complexity by transitioning from a flat, two-dimensional view of neural networks to a multi-dimensional hierarchical structure with multiple layers stacked vertically. This dimensional change allows the system to organize neurons and connections in three-dimensional space, enabling efficient routing and reduced interconnection complexity. By adding the layer dimension, the patent can manage enormous numbers of parameters through structured organization rather than random connectivity, reducing the practical complexity of implementation.