@TomAndJerry that's a good question about patents and one I asked myself. This is a great new piece of tech and it needs protecting. Would need to confirm with the company, but from what I've read, I think it falls under patent AU2023214246A1, Title EVENT-BASED EXTRACTION OF FEATURES IN A CONVOLUTIONAL SPIKING NEURAL NETWORK, specifically:
[0130] The conventional unsupervised learning rule finds repeating patterns in data. It is robust to noise and often only requires the pattern to be repeated 2-5 times. The conventional rule can perform dimensionality reduction because neurons learn to become sensitive to only a single pattern and therefore only send a few events indicating whether a particular pattern has been detected. The conventional learning rule detects patterns because the conventional learning rule can only detect a pattern within a packet, which is a collection of events that are close to one another in time. The conventional learning rule assumes binary synapses (weights) that can be either 0 or 1 and binary or ternary activations that can have values of either 0, 1, and in the ternary case, -1. The descriptions will consider the binary weights/binary activations case.
[0229] Other operations that transform the input data in order to produce modified versions of the input data for training data set augmentation purposes are also envisaged. For example, a transformation operation may be applied wherein selected data present in the input data is obscured, removed or modified, for example in order to simulate noise or distortion effects.
[0230] In one such transformation operation, the data in the input is masked or relevant input neurons are masked, controlled or otherwise arranged to prevent the data from being incorporated in the subsequent convolution operations. For example, selected input coordinates may be ignored, data associated with selected input coordinates may be zeroed, or new data could be added to an input coordinate, for example randomly, to simulate random noise, and/or to simulate particular color space manipulations by masking color channels. Such data changing operations may be carried out in addition to
any one or more other transformation operation.
[0234] In some embodiments,
noise is introduced to an image by randomly masking and/or carrying out modification actions at particular input layer neurons.
---------------------------------------------------------------------------------------------------------------------------------------------------------------
The patent is 85 pages long, so I used GPT4o to cross reference to TENNS:
The excerpt from patent AU2023214246A1, titled EVENT-BASED EXTRACTION OF FEATURES IN A CONVOLUTIONAL SPIKING NEURAL NETWORK, seems to describe a system for recognizing patterns in data using a spiking neural network (SNN). This includes processes like:
- Noise handling: Introducing or simulating noise in the input data, masking specific parts of the data to improve robustness in learning.
- Pattern recognition: Using neurons to detect recurring patterns in input data.
- Dimensionality reduction: Learning to become sensitive to specific patterns, thus reducing the complexity of data representation.
Relation to TENNS and aTENNuate:
TENNS (Temporal Event-Based Neural Network System):
- TENNS is designed for event-based processing, focusing on recognizing and processing patterns over time, which aligns well with the convolutional spiking neural network (CSNN) described in this patent.
- The mention of dimensionality reduction and learning patterns over repeated events (like 2-5 times) could directly relate to the temporal and event-based nature of TENNS.
aTENNuate (Noise Clarification Tool):
- aTENNuate deals with real-time noise removal and enhancing speech clarity, which is closely related to the noise handling section in this patent. The process of introducing and masking noise, and simulating distortion effects, could be part of the techniques used in aTENNuate to train the model for dealing with noisy input (speech).
- The patent's focus on masking parts of the input data to simulate noise directly connects to what aTENNuate does in real time for speech signals.
Conclusion:
The concepts described in the excerpt from patent AU2023214246A1 are relevant to both TENNS and aTENNuate:
- For TENNS, the event-based processing and learning rules resonate with its temporal pattern recognition approach.
- For aTENNuate, the noise simulation and masking operations closely mirror the noise reduction functionality.
Therefore, this patent could cover aspects of the underlying technology for both tools.
So, it would appear, at least on the surface, that the technology is covered by this patent, which was filed approximately 12 months ago, aligning well with the expected development timeline. Happy dayz.
(20min delay)