US 10,984,320 Highly trainable neural network configuration
5 defendants · 5 suits · updated September 14, 2026
Also cited as U.S. Patent No. 10,984,320 · US10984320B2 · US 10,984,320 B2 · 10984320 · the '320 patent
Tesseract Systems LLC has filed 5 patent-infringement suits on US 10,984,320, "Highly trainable neural network configuration" since May 27, 2026, all in the D. Del., against 5 distinct defendants including AssemblyAI Inc., Cerence Inc. and Deepgram, Inc.. The patent was granted April 20, 2021 on a priority date of May 2, 2016. Its abstract opens: "A computer-based method includes receiving an input signal at a neuron in a computer-based neural network that includes a plurality of neuron layers, applying a first non-linear transform to the input signal at the neuron to produce a plain signal, and…" The examiner cited 3 references during prosecution.
- Assignee
- Tesseract Systems LLC
- Priority date
- May 2, 2016
- Granted
- April 20, 2021
- Expires
- May 2, 2036
- Status
- Active
- Claims
- 28 (4 independent)
- Inventors
- Rupesh Kumar Srivastava, Klaus Greff
- Full text
- Google Patents
The suits
5 cases · D. Del. (5)
| Filed | Defendant | Plaintiff | Court | Docket |
|---|---|---|---|---|
| 2026-05-27 | AssemblyAI Inc. | Tesseract Systems LLC | D. Del. | 1:26-cv-00605 |
| 2026-05-27 | Cerence Inc. | Tesseract Systems LLC | D. Del. | 1:26-cv-00606 |
| 2026-05-27 | Deepgram, Inc. | Tesseract Systems LLC | D. Del. | 1:26-cv-00607 |
| 2026-05-27 | Five9, Inc. | Tesseract Systems LLC | D. Del. | 1:26-cv-00609 |
| 2026-05-27 | SoundHound, Inc. | Tesseract Systems LLC | D. Del. | 1:26-cv-00611 |
Docket links open the public record on CourtListener. Filing dates are the complaint's.
Independent claims
What is actually asserted, verbatim.
1. A computer-based method of facilitating training in a computer-based neural network, the method comprising: receiving an electrical input signal at a neuron in the computer-based neural network that comprises a plurality of neuron layers; applying a first non-linear transform to the input signal at the neuron to produce a plain signal; applying a second non-linear transform to the input signal at a first gate in the neuron to produce a transform signal; applying a third non-linear transform to the input signal at a second gate in the neuron to produce a carry signal; and calculating a weighted sum of a non-transformed first component of the input signal and the plain signal at the neuron, wherein calculating the weighted sum of the non-transformed first component of the input signal and the plain signal comprises: calculating, with at least one computer-based multiplier, a product of the plain signal and the transform signal to produce a weighted plain signal; calculating, with the at least one computer-based multiplier, a product of the non-transformed first component of the input signal and the carry signal to produce a weighted input signal component; and adding the weighted plain signal to the weighted input signal component to produce the weighted sum.
15. A computer-based neural network comprising: a plurality of layers of neurons, wherein each layer of neurons comprises a plurality of neurons, and wherein each of one or more of the neurons: comprises an input to receive an electrical input signal; applies a first non-linear transform to the input signal to produce a plain signal; applies, with a first gate of the neuron, a second non-linear transform to the input signal to produce a transform signal; and applies, with a second gate of the neuron, a third non-linear transform to the input signal to produce a carry signal, calculates a weighted sum of a non-transformed first component of the input signal and the plain signal; and comprises an output to send the calculated weighted sum out of the neuron, and wherein the computer-based neural network and each of the neurons is implemented via a processor executing instructions stored on a computer-readable medium, wherein calculating the weighted sum of the non-transformed first component of the input signal and the plain signal is performed by the processor executing the instructions stored on the computer-readable medium embodying: at least one multiplier module configured to: calculate a product of the plain signal and the transform signal to produce a weighted plain signal; and calculate a product of the non-transformed first component of the input signal and the carry signal to produce a weighted input signal component; and an adder module configured to add the weighted plain signal to the weighted input signal component to produce the weighted sum, wherein the at least one multiplier module and the adder module are implemented via the processor executing instructions stored on the computer-readable medium.
26. A non-transitory, computer-readable medium that stores instructions executable by a processor to perform the steps to facilitate training of a computer-based neural network, the steps comprising: receiving an electrical input signal at a neuron in the computer-based neural network that comprises a plurality of neuron layers; applying a first non-linear transform to the input signal in the neuron to produce a plain signal; applying a second non-linear transform to the input signal at a first gate in the neuron to produce a transform signal; applying a third non-linear transform to the input signal at a second gate in the neuron to produce a carry signal; and calculating a weighted sum of a non-transformed first component of the input signal and the plain signal at the neuron, wherein calculating the weighted sum of the non-transformed first component of the input signal and the plain signal at the neuron comprises: calculating, with at least one computer-based multiplier, a product of the plain signal and the transform signal to produce a weighted plain signal; calculating, with the at least one computer-based multiplier, a product of the non-transformed first component of the input signal and the carry signal to produce a weighted input signal component; and adding, with a computer-based adder, the weighted plain signal to the weighted input signal component, to produce the weighted sum.
28. A method comprising: providing a computer-based neural network comprising a plurality of layers of neurons, wherein each layer of neurons comprises a plurality of neurons, and wherein each respective one of one or more of the neurons comprises: an input to receive an electrical input signal; and an output to send an electrical output signal to every one of the neurons in a subsequent one of the neuron layers; wherein the computer-based neural network and each of the neurons in the computer-based neural network is implemented via a processor executing instructions stored on a computer-readable medium, receiving an electrical input signal at a first neuron in a first layer of the computer-based neural network; applying a first non-linear transform to the input signal at the first neuron to produce a plain signal; applying a second non-linear transform to the input signal at a first gate in the first neuron to produce a transform signal; applying a third non-linear transform to the input signal at a second gate in the first neuron to produce a carry signal; and calculating a weighted sum of a non-transformed first component of the input signal and the plain signal at the first neuron, wherein calculating the weighted sum of the non-transformed first component of the input signal and the plain signal comprises: calculating, with at least one computer-based multiplier, a product of the plain signal and the transform signal to produce a weighted plain signal; calculating, with the at least one computer-based multiplier, a product of the non-transformed first component of the input signal and the carry signal to produce a weighted input signal component; and adding, with a computer-based adder, the weighted plain signal to the weighted input signal component to produce the weighted sum, wherein calculating the weighted sum of the non-transformed first component of the input signal and the plain signal is performed by the processor executing the instructions stored on the computer-readable medium embodying the computer-based multiplier and the computer-based adder, and sending an electrical signal representing the weighted sum of the non-transformed first component of the input signal and the plain signal out from the output of the first neuron of the computer-based neural network to each respective one of a plurality of neurons in a second layer of neurons of the computer-based neural network; and adjusting one or more parameters associated with the first non-linear transform, and/or the second non-linear transform, and/or the third non-linear transform, at each of one or more of the neurons.
Art before the examiner
16 references of record, 3 cited by the examiner.
Everything listed here was before the Office and is what a challenge must go beyond. Art the examiner never saw is where an invalidity search starts.
Examiner-cited
- US20020184174A1 — Finite-state automaton modeling biologic neuron (Kadri Rachid M.)
- US7904398B1 — Artificial synapse component using multiple distinct learning means with distinct predetermined learning acquisition times (Dominic John Repici)
- US20170185897A1 — Self-stabilized deep neural network (Microsoft Technology Licensing, Llc)
Applicant-cited and other (13)
- US5214715A — Predictive self-organizing neural network
- US5742741A — Reconfigurable neural network
- US5864803A — Signal processing and training by a neural network for phoneme recognition
- US6882992B1 — Neural networks for intelligent control
- US7321882B2 — Method for supervised teaching of a recurrent artificial neural network
- US7395521B1 — Method and apparatus for translating an imperative programming language description of a circuit into a hardware description
- US7930260B2 — System and method for real time pattern identification
- US8103606B2 — Architecture, system and method for artificial neural network implementation
- US8311958B2 — Hierarchical temporal memory methods and systems
- US8775341B1 — Intelligent control with hierarchical stacked neural networks
- US8943008B2 — Apparatus and methods for reinforcement learning in artificial neural networks
- US20150364127A1 — Advanced recurrent neural network based letter-to-sound
- US20150364128A1 — Hyper-structure recurrent neural networks for text-to-speech
Classification
- G06N3/084primary — Backpropagation, e.g. using gradient descent
- G06N3/048 — Activation functions
- G06N3/0481
Family
- US10984320B2 — B2 (2021)
- US20170316308A1 — A1 (2017)
Abstract
A computer-based method includes receiving an input signal at a neuron in a computer-based neural network that includes a plurality of neuron layers, applying a first non-linear transform to the input signal at the neuron to produce a plain signal, and calculating a weighted sum of a first component of the input signal and the plain signal at the neuron. In a typical implementation, the first non-linear transform is a function of the first component of the input signal and at least a second component of the input signal.
Named in one of these suits? A prior-art search is commissioned once and serves every defendant who shares it. The examiner's own search record is free on the tools page, and your first preliminary search is free — one claim, back within 24 hours — so you know whether there is art the Office never saw before anyone pays for more. Flat fees: $2,500$899 for the independent claims, $5,000$1,500 for every asserted claim.
This page is assembled from public records — federal court dockets and the patent's own file — and is updated as the case list changes. It states what the record shows and is not an opinion on the validity or infringement of any claim. Patent data via Google Patents; docket data via CourtListener.