Tesseract Systems LLC v. Deepgram Inc

I. Executive Summary and Procedural Information

• Parties & Counsel:
  • Plaintiff: Tesseract Systems LLC (New Mexico)
  • Defendant: Deepgram, Inc. (Delaware)
  • Plaintiff's Counsel: Silverman, McDonald & Friedman
• Case Identification: 1:26-cv-00607, D. Del., 05/27/2026
• Venue Allegations: Venue is alleged to be proper in the District of Delaware because the Defendant has an established place of business in the district, has committed acts of patent infringement in the district, and has caused harm to the Plaintiff in the district.
• Core Dispute: Plaintiff alleges that Defendant's neural network products and services infringe a patent related to a specific "highway network" architecture designed to improve the trainability of deep neural networks.
• Technical Context: The technology addresses the challenge of training very deep artificial intelligence models, a fundamental problem in machine learning where performance can degrade as network complexity increases.
• Key Procedural History: The complaint does not mention any prior litigation, Inter Partes Review (IPR) proceedings, or licensing history related to the patent-in-suit.

Case Timeline

Date	Event
2016-05-02	'320' Patent Priority Date
2017-05-01	'320 Patent Application Filing Date
2021-04-20	'320 Patent Issue Date
2026-05-27	Complaint Filing Date

II. Technology and Patent(s)-in-Suit Analysis

Patent Identification

• Patent Identification: U.S. Patent No. 10,984,320, "Highly trainable neural network configuration," issued April 20, 2021.

The Invention Explained

• Problem Addressed: The patent's background section describes the difficulty of training deep neural networks, noting that simply adding more layers makes optimization "considerably more difficult" '320 Patent, col. 1:36-39 Deeper "plain" networks can suffer from issues like signal degradation that impede the learning process '320 Patent, Fig. 7
• The Patented Solution: The invention proposes a "highway network" architecture that introduces a "learned gating mechanism for regulating information flow" '320 Patent, col. 2:34-36 As illustrated in Figure 1, each "highway neuron" contains gates that determine whether to transform an input signal or to carry it through to the next layer unaltered '320 Patent, Fig. 1 '320 Patent, col. 5:40-54 This creates "information highways" that allow data to flow across many layers "without attenuation," which in turn makes it possible to effectively train networks of "virtually arbitrary depth" using standard optimization methods '320 Patent, col. 2:32-44
• Technical Importance: This architecture enabled the effective training of significantly deeper neural networks (e.g., up to 900 layers) than was previously practical, facilitating research into the impact of network depth on complex problems '320 Patent, col. 2:42-44 '320 Patent, col. 2:60-63

Key Claims at a Glance

• The complaint asserts "Exemplary '320 Patent Claims" but does not identify specific claims, instead referencing an external exhibit not provided with the complaint Compl. ¶11 The first independent claim, Claim 1, is representative of the patented method.
• The essential elements of independent Claim 1 include:
  • Receiving an electrical input signal at a neuron in a neural network.
  • Applying a first non-linear transform to the input to produce a "plain signal."
  • Applying a second non-linear transform at a first gate to produce a "transform signal."
  • Applying a third non-linear transform at a second gate to produce a "carry signal."
  • Calculating a weighted sum of a "non-transformed first component of the input signal" and the "plain signal," where this calculation involves multiplying the plain signal by the transform signal and multiplying the non-transformed input component by the carry signal, then adding the two products.
• The complaint does not explicitly reserve the right to assert dependent claims, but refers generally to infringement of "one or more claims" Compl. ¶11

III. The Accused Instrumentality

Product Identification

• The complaint refers to "Exemplary Defendant Products" without naming them specifically Compl. ¶11 It states that these products are identified in "Exhibit 2," which contains claim charts but was not filed with the public complaint Compl. ¶16 Compl. ¶17

Functionality and Market Context

• The complaint provides no specific details regarding the functionality of the accused products beyond alleging they "practice the technology claimed by the '320 Patent" Compl. ¶16 Defendant Deepgram, Inc. is known to operate in the field of artificial intelligence and speech recognition. The complaint alleges these products are made, used, sold, and imported by the Defendant Compl. ¶11 No probative visual evidence provided in complaint.

IV. Analysis of Infringement Allegations

The complaint alleges that infringement is detailed in claim charts provided as Exhibit 2 Compl. ¶16 As this exhibit was not provided, a claim chart summary cannot be constructed. The complaint's narrative theory is that the "Exemplary Defendant Products" practice the patented technology and therefore "satisfy all elements of the Exemplary '320 Patent Claims" Compl. ¶16 Infringement is alleged to occur through Defendant's making, using, and selling of the products, as well as through internal testing by its employees Compl. ¶¶11-12

Identified Points of Contention

• Scope Questions: The claims require a specific gating structure involving a "transform signal" and a "carry signal" produced by distinct non-linear transforms '320 Patent, cl. 1 A central question may be whether modern neural network architectures, such as those based on residual connections (ResNets) or other advanced designs, can be mapped onto this specific "highway network" formulation, or if they represent a distinct, non-infringing approach to solving the deep network training problem.
• Technical Questions: Claim 1 recites calculating a weighted sum using a "non-transformed first component of the input signal" '320 Patent, cl. 1 The infringement analysis may turn on what evidence Plaintiff can provide to show that the accused products use a "component" of the original, untransformed input in this specific manner, as opposed to using the entire input vector or a different value in their calculations.

V. Key Claim Terms for Construction

The Term: "non-transformed first component of the input signal"

• Context and Importance: This term appears in the final "calculating a weighted sum" step of Claim 1 and is critical for defining the precise mathematical operation at the core of the invention. Whether Defendant's products infringe may depend heavily on whether their architecture uses an untransformed "component" of the input or the entire input vector in the summing step.
• Intrinsic Evidence for Interpretation:
  • Evidence for a Broader Interpretation: The patent includes a general formula for the neuron's output: y=H(x, WH) * T(x, W₁) + x * (1-T(x, W₁)) '320 Patent, col. 11:40-42 In this formula, x represents the entire input vector, not just a component, which could support an interpretation where "component" can encompass the whole input.
  • Evidence for a Narrower Interpretation: The claim language itself specifies a "first component," suggesting a portion rather than the whole. More pointedly, a detailed numerical example in the specification shows the calculation for a single neuron's output yi as yi = Hi(x)Ti(x) + xiCi(x), where xi is explicitly defined as "the i-th element of vector x" '320 Patent, col. 14:30-34 This provides strong intrinsic evidence for a narrower construction where the "component" is a single corresponding element of the input vector.

The Term: "plain signal"

• Context and Importance: The "plain signal" is the output of the "first non-linear transform" and an input to the final weighted sum calculation '320 Patent, cl. 1 Its definition is foundational to the claim, as it is the transformed data path that is gated against the untransformed "carry" path.
• Intrinsic Evidence for Interpretation:
  • Evidence for a Broader Interpretation: "Plain signal" is not a standard term of art. A party might argue it simply refers to the output of the main transformation H(x, WH) before any gating is applied, without further limitation on the nature of H '320 Patent, col. 11:23-24
  • Evidence for a Narrower Interpretation: The term is used to distinguish this signal from the "transform signal" and "carry signal," which are generated by gates T and C '320 Patent, cl. 1 A party could argue that the term "plain" implies a non-gated transformation, creating a structural distinction that must be present in an accused device. Figure 1 visually separates the transform H from the gates T and C, which may support this narrower view '320 Patent, Fig. 1

VI. Other Allegations

Indirect Infringement

• The complaint alleges induced infringement based on Defendant's sale of the accused products combined with the distribution of "product literature and website materials" that allegedly instruct customers on how to use the products in an infringing manner Compl. ¶14 Compl. ¶15 This allegation is directed at conduct occurring at least since the filing of the complaint Compl. ¶15

Willful Infringement

• The basis for willfulness is post-suit knowledge. The complaint alleges that service of the complaint itself provides Defendant with "Actual Knowledge of Infringement" and that any continued infringing activity thereafter is willful Compl. ¶13 Compl. ¶14 No allegations of pre-suit knowledge are made.

VII. Analyst's Conclusion: Key Questions for the Case

• A core issue will be one of definitional scope: can the claim term "non-transformed first component of the input signal" be construed to cover the entire input vector, as suggested by a general formula in the patent '320 Patent, col. 11:40-42, or is it limited to a single corresponding element of the input, as explicitly described in the patent's detailed numerical example '320 Patent, col. 14:30-34?

• A second key issue will be one of technical mapping: will the Plaintiff be able to demonstrate that the architecture of Defendant's modern AI models contains the distinct structural elements of a "highway network," including the specific "plain signal," "transform signal," and "carry signal" pathways, or will Defendant successfully argue that its products use a fundamentally different, non-infringing architecture such as a standard residual network (ResNet)?

• Finally, a primary threshold question will be an evidentiary one: because the complaint's infringement allegations rely entirely on an unprovided exhibit Compl. ¶16, the case may initially hinge on what specific technical evidence Plaintiff presents to show that Defendant's products perform each step of the claimed method, especially the precise gating and weighted sum calculations required by Claim 1.