BrainChip has acquired an Australian patent for “An Improved Spiking Neural Community,” which outlines a method for enhancing the educational capabilities of spiking neural networks utilizing low-shot studying.
The approach permits the community to study new duties with fewer coaching examples, which is especially advantageous in conditions the place knowledge is dear to accumulate.
The patent additionally mentions the addition of latest neurons to the ultimate layer of a beforehand educated community. Considerably, solely the weights of those new neurons require coaching with a number of new samples, whereas the rest of the community stays unchanged. This protects the computational sources.
“BrainChip’s increasing patent portfolio ensures our freedom to observe our personal innovations and to stop others from infringing on our extremely progressive mental property. This newest Australian patent is the subsequent step in advancing edge AI additional than ever earlier than,” says Sean Hehir, CEO of BrainChip.
BrainChip’s claims doc outlines a state-of-the-art system that includes a neural processor. This processor can obtain modification requests to reinforce a base neural community. Particularly, this modification entails including a number of spiking neurons to the Nth layer of the bottom neural community. A number of the unique spiking neurons are retained in the identical layer.
The system additionally has a neuron material incorporating the neural processor and a reminiscence. This reminiscence shops the neural community configuration, which defines the connections between spiking neurons and synapses. It additionally consists of details about their membrane potential values and synaptic weights.
The bottom neural community’s studying operate considers the membrane potential values of each spiking neuron in a given layer. Within the modified community, the choice course of for a successful spiking neuron excludes the membrane potential worth of at the least one neuron within the Nth layer. The neural processor then identifies the successful spiking neuron from the supplementary spiking neurons by deciding on the one with the best membrane potential worth.
The patent has implications throughout numerous purposes, together with biometric face recognition, speech recognition, and anomaly detection in industrial methods. BrainChip has over 20 issued patents and 23 pending patent purposes.
Associated
Brainchip | edge AI | neural networks | patent