Feed-forward spike-based neural network
“Neuromorphic” computing systems use electronic circuits integrated on silicon chips to mimic biological neural networks and do robust, low-power computation. Neurons are typically arranged in one or two-dimensional arrays, and like in a biological system require dense connections between them, getting computational power from their parallelism. At times this is achieved using wires, a shared routing infrastructure or stacking the arrays and using metal interconnects between the layers. Alternatively, a digitally encoded number can be transmitted which represents the “address” of the neuron, so that the neuron circuit which produced an event can be uniquely identified. Nevertheless, the communication in these architectures remains a bottleneck, compared to the native 3D connectivity in the brain, contributing overwhelmingly to the power budget of neuromorphic computing systems.
A layered spike-based neural network is described that overcomes the aforementioned shortfall of current state of the art. A simple 2D printing process uses thin-film transistors to create flexible PCBs. These are then coiled, and transmission (preferentially pulsed capacitive transmission) between tightly-packed layers allow neural networks to be formed with 3-dimensional wiring, giving the structure advantages for power consumption and architectural simplicity and the ability to be integrated with various forms of sensors, in particular in combination with patent application IT202200000779. Various augmentations to and variations of the basic design can be developed.
- Highly sensitive and perceptive robotic tactile sensors (incl. industrial, household, surgical, prosthetics);
- Biosensors (incl. wearable, ingestible);
- Biodegradable multimodal environmental monitoring (incl. agricultural, soil, forestry, water);
- Multimodal perception (chemical, thermal, tactile, optical, hyperspectral etc).
- Simple and economical method – reel-to-reel printing offers pathway to commercialisation;
- Computational power scalability advantage over state-of-the-art neuromorphic systems;
- Resulting devices can be completely biodegradable;
- Integration with sensing and actuation for complete nervous systems from single printing process.