BrainChip patents for improving the learning capabilities of spiking neural networks
BrainChip has received an Australian patent for “An Improved Spiking Neural Network,” which outlines a strategy for enhancing the learning capabilities of spiking neural networks using low-shot learning.
The technique allows the network to learn new tasks with fewer training examples, which is particularly advantageous in situations where data is expensive to acquire.
The patent also mentions the addition of new neurons to the final layer of a previously trained network. Significantly, only the weights of these new neurons require training with a few new samples, while the remainder of the network remains unchanged. This saves the computational resources.
“BrainChip’s expanding patent portfolio ensures our freedom to practice our own inventions and to prevent others from infringing on our highly innovative intellectual property. This latest Australian patent is the next step in advancing edge AI further than ever before,” says Sean Hehir, CEO of BrainChip.
BrainChip’s claims document outlines a state-of-the-art system featuring a neural processor. This processor can receive modification requests to enhance a base neural network. Specifically, this modification involves adding one or more spiking neurons to the Nth layer of the base neural network. Some of the original spiking neurons are retained in the same layer.
The system also has a neuron fabric incorporating the neural processor and a memory. This memory stores the neural network configuration, which defines the connections between spiking neurons and synapses. It also includes information about their membrane potential values and synaptic weights.
The base neural network’s learning function considers the membrane potential values of every spiking neuron in a given layer. In the modified network, the selection process for a winning spiking neuron excludes the membrane potential value of at least one neuron in the Nth layer. The neural processor then identifies the winning spiking neuron from the supplementary spiking neurons by selecting the one with the highest membrane potential value.
The patent has implications across various applications, including biometric face recognition, speech recognition, and anomaly detection in industrial systems. BrainChip has over 20 issued patents and 23 pending patent applications.
Comments