Artificial intelligence framework combining a spiking neural network and a hyperdimensional computing block
View Patent ↗An artificial intelligence framework is disclosed having a spiking neural network configured to extract low-level features from event-based spiking data and provide the low-level features as a spiking neural network output signal. Further included is a hyperdimensional computing block that is configured to map the spiking neural network output into high-dimensional space and classify data from the abstract information.
1 . An artificial intelligence framework comprising:
a spiking neural network (SNN) implemented in electronic hardware configured to extract low-level features from event-based spiking data and provide the low-level features as a spiking neural network output signal; and
a hyperdimensional computing (HDC) block implemented in an electronic hardware block comprised of logic gates and an HDC memory, wherein the HDC block comprises:
an HDC encoder configured to map the spiking neural network output signal into high-dimensional space and classify data from the abstract information; and
an HDC learning model that is configured to combine encoded hypervectors and generate a hypervector for each class within a set of associated classes.
2 . The artificial intelligence framework of claim 1 wherein the HDC encoder is a non-linear neural encoder.
3 . The artificial intelligence framework of claim 1 wherein the HDC learning model is configured to store the encoded hypervectors within the HDC memory.
4 . The artificial intelligence framework of claim 1 wherein the SNN is configured to be trained based on a prediction made by the HDC learning model.
5 . The artificial intelligence framework of claim 1 wherein the HDC block is configured to perform a Moore-Penrose inverse of HDC encoding performed by the HDC encoder.
6 . The artificial intelligence framework of claim 5 wherein the HDC block is configured to apply the Moore-Penrose inverse of HDC encoding to back-propagate loss from the HDC learning model to the SNN.
7 . The artificial intelligence framework of claim 1 wherein the HDC encoder is configured to use a hyperbolic tangent function as an activation function.
8 . The artificial intelligence framework of claim 1 wherein the HDC encoder is configured with an encoding method that employs a kernel trick to map data points into the high-dimensional space.
9 . The artificial intelligence framework of claim 8 wherein the kernel trick is applied to a Gaussian kernel.
10 . The artificial intelligence framework of claim 1 wherein the SNN is configured to pass a training datum to the HDC encoder.
11 . The artificial intelligence framework of claim 10 wherein the HDC block is encoded to add each encoded training datum to class hypervectors depending on how much new information an identified pattern adds to the class hypervectors.
12 . The artificial intelligence framework of claim 1 wherein the SNN is configured to update using a gradient descent rule.
13 . The artificial intelligence framework of claim 12 wherein the HDC block is configured to achieve single pass learning.
14 . The artificial intelligence framework of claim 1 wherein the SNN is configured to have only a single digit number of SNN layers to extract low-level spatiotemporal information of raw event-based data.
15 . The artificial intelligence framework of claim 1 further configured to support online learning from an Internet data stream.
16 . The artificial intelligence framework of claim 1 wherein the SNN and the HDC block are configured to be co-trained.
17 . The artificial intelligence framework of claim 1 wherein the HDC block is configured to be trained based on a loss function defined at an output layer of the SNN.
18 . The artificial intelligence framework of claim 1 wherein the electronic hardware block comprises a field-programmable gate array.