DCT

1:25-cv-01588

Health Discovery Corp v. NVIDIA Corp

Log In
Key Events
Complaint

I. Executive Summary and Procedural Information

  • Parties & Counsel:
  • Case Identification: 1:25-cv-01588, D. Del., 12/31/2025
  • Venue Allegations: Plaintiff asserts venue is proper in the District of Delaware because Defendant is a Delaware corporation, is subject to personal jurisdiction, has previously litigated in the district without challenging venue, and has allegedly committed acts of patent infringement in the district.
  • Core Dispute: Plaintiff alleges that Defendant’s GPU-accelerated machine learning software stack infringes a patent related to a method for improving the performance of Support Vector Machine (SVM) classifiers by recursively eliminating irrelevant or redundant data features.
  • Technical Context: The technology addresses the field of machine learning, specifically improving the efficiency and accuracy of data classification systems by automating the selection of the most informative features from high-dimensional datasets.
  • Key Procedural History: The complaint notes that the patent-in-suit, U.S. Patent No. 10,402,685, survived an inter partes review challenge when the Patent Trial and Appeal Board declined to institute proceedings in 2021. The patent was also previously asserted against Intel Corporation in litigation that began in 2020 and was resolved in September 2023.

Case Timeline

Date Event
1999-10-27 ’685 Patent Priority Date
2019-09-03 ’685 Patent Issue Date
2020-07-23 Plaintiff files infringement action against Intel Corp.
2021-02-01 Intel Corp. requests inter partes review of the ’685 Patent
2021-01-01 PTAB declines to institute inter partes review of the ’685 Patent
2022-04-04 Plaintiff files second infringement action against Intel Corp.
2022-05-26 Federal Circuit remands the 2020 Intel action
2023-09-01 Plaintiff and Intel Corp. resolve disputes
2025-12-31 Complaint filed against NVIDIA Corp

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

U.S. Patent No. 10,402,685 - *“Recursive Feature Elimination Method Using Support Vector Machines”*

The Invention Explained

  • Problem Addressed: The patent addresses the challenge of "feature selection" in machine learning, where classifiers like Support Vector Machines (SVMs) are trained on datasets with a very large number of features (e.g., thousands of genes for dozens of patients) (Compl. ¶27; ’685 Patent, col. 2:37-49). This high dimensionality can lead to "overfitting," where a model performs well on training data but fails to generalize to new data, and can also create an impractical computational burden (’685 Patent, col. 2:34-39, 2:63-3:6).
  • The Patented Solution: The invention discloses a method termed Recursive Feature Elimination (RFE) specifically for SVMs (Compl. ¶27). Instead of selecting features based on simple correlation, the method uses the SVM's own internal logic. It first trains the SVM, then uses the weights assigned by the classifier to its input features to compute a ranking criterion for those features. It then eliminates the feature with the smallest ranking criterion and repeats the process. This iterative pruning yields an optimized subset of features that improves the classifier's efficiency and predictive power (’685 Patent, Abstract; col. 17:39-49).
  • Technical Importance: The claimed process automates feature selection, transforming it from a manual or heuristic task into a mathematically driven optimization loop integrated within the learning process itself (Compl. ¶19).

Key Claims at a Glance

  • The complaint asserts infringement of at least method claim 1 (Compl. ¶34).
  • The essential elements of independent claim 1 include:
    • Retrieving training data for a processor.
    • The processor determining a value for each feature in the data.
    • Eliminating at least one feature that has a "minimum ranking criterion," where that criterion is based on the value determined for each feature.
    • "Subtracting a matrix from the kernel data" to create an "updated kernel data," with the matrix being derived from training samples corresponding to the eliminated feature.
    • Updating the value for each remaining feature based on the updated kernel data.
    • Repeating the elimination and updating steps until a predetermined number of features remains, which generates a feature ranking list.
    • Using the resulting ranked list to recognize new data.
  • The complaint does not explicitly reserve the right to assert dependent claims but notes its allegations are "exemplary only" (Compl. p. 21, n.2).

III. The Accused Instrumentality

Product Identification

  • Defendant’s GPU-accelerated machine learning stack, which includes RAPIDS cuML and the associated cuml.accel module (Compl. ¶34).

Functionality and Market Context

  • The complaint alleges that the accused products are designed to execute machine learning pipelines, such as those built with the popular scikit-learn library, on Nvidia GPUs (Compl. ¶34). Specifically, Plaintiff identifies workflows named "RFE(SVC(kernel="linear"))" and "RFECV(SVC(kernel="linear"))" that Nvidia allegedly "publishes, promotes, and instructs developers to use" (Compl. ¶34). These workflows are alleged to implement an iterative process of training an SVM, ranking features, eliminating the lowest-ranked features, and updating the model, thereby mirroring the patented method (Compl. ¶¶20, 34).

IV. Analysis of Infringement Allegations

The complaint alleges that when developers use the accused RAPIDS cuML software as instructed by Nvidia, they perform each step of the patented method (Compl. ¶34). The complaint reproduces Figure 2 of the '685 patent, a flowchart that illustrates the iterative training and kernel adjustment process at the core of the invention (Compl. p. 9, ¶26).

’685 Patent Infringement Allegations

Claim Element (from Independent Claim 1) Alleged Infringing Functionality Complaint Citation Patent Citation
retrieving training data from one or more storage devices in communication with a processor, The accused workflows retrieve training data to be processed on Nvidia GPUs. ¶34 col. 8:51-52
the processor operable for: determining a value for each feature in a group of features provided by the training data; The underlying SVM in the accused workflows computes per-feature values (weights) during the model training phase. ¶34 col. 19:22-26
eliminating at least one feature with a minimum ranking criterion from the group, The accused workflows eliminate the lowest-ranked features based on the values computed during SVM training. ¶34 col. 19:35-37
wherein the minimum ranking criterion is obtained based on the value for each feature in the group; The ranking criterion is derived from the feature weights calculated by the SVM. ¶34 col. 19:33-34
subtracting a matrix from the kernel data to provide an updated kernel data, The accused workflows update kernel relationships through a "refitting" process after a feature is eliminated. ¶34 col. 21:3-11
each component of the matrix comprising a dot product of two of training samples...that corresponds to the eliminated feature; The alleged matrix subtraction mathematically alters the kernel representation of the SVM, which is based on dot products of training data. ¶36 col. 21:3-11
updating the value for each feature of the group based on the updated kernel data; The "refitting" process in the accused workflows recomputes an updated kernel matrix with the eliminated features removed. ¶34 col. 19:15-17
repeating of eliminating the at least one feature from the group and updating the value for each feature of the group until a number of features in the group reaches a predetermined value to generate a feature ranking list; and The accused workflows repeat the elimination-and-refit loop until a specified feature count is reached. ¶34 col. 19:15-17
recognizing a new data corresponding to the group of features with the feature ranking list. The final, feature-ranked SVM is used to classify new data. ¶34 col. 9:11-17

Identified Points of Contention

  • Scope Questions: The primary point of contention may be the claim element "subtracting a matrix from the kernel data to provide an updated kernel data." The complaint alleges this is met by "updating kernel/Gram relationships through refitting" (Compl. ¶34). The dispute may turn on whether the accused "refitting" process, which involves retraining the model on the reduced feature set, is equivalent to the specific mathematical operation of "subtracting a matrix" as required by the claim.
  • Technical Questions: A key question for the court will be whether the internal computations performed by Nvidia's cuML SVM implementations during the RFE and RFECV workflows perform the specific steps recited in the claim. The analysis will require evidence showing how the accused software mathematically modifies the SVM's kernel representation after each feature elimination step.

V. Key Claim Terms for Construction

  • The Term: "subtracting a matrix from the kernel data"
  • Context and Importance: This term is central to the dispute because it describes the specific mechanism by which the claimed invention modifies the SVM classifier during the iterative process. The infringement allegation hinges on whether the accused product's "refitting" operation falls within the scope of this term. Practitioners may focus on this term because the complaint’s infringement theory appears to equate a general "refitting" with a specific mathematical operation.
  • Intrinsic Evidence for Interpretation:
    • Evidence for a Broader Interpretation: A party might argue that the term should be interpreted functionally to cover any process that mathematically updates the kernel representation to reflect the removal of a feature. The patent's specification describes a more complex process of recomputing the kernel matrix H to create H(-i) after removing feature i, which could be argued to be conceptually equivalent to, if not literally, "subtracting a matrix" (’685 Patent, col. 21:12-22).
    • Evidence for a Narrower Interpretation: A party could argue that the plain language of the claim requires a literal matrix subtraction operation, not a more general "refitting" or full recalculation of the kernel. The claim recites "subtracting a matrix," not "recomputing," "refitting," or "updating" the kernel data. An argument could be made that the patentee chose this specific language to distinguish the invention from prior art methods that might have involved complete retraining at each step.

VI. Other Allegations

  • Indirect Infringement: The complaint alleges facts that may support a claim for induced infringement by stating that Nvidia "publishes, promotes, and instructs developers to use" the accused workflows, thereby encouraging them to perform the steps of the patented method (Compl. ¶34).
  • Willful Infringement: The complaint alleges that Defendant’s infringement has been "willful, deliberate, and intentional" (Compl. ¶66). The complaint does not, however, plead specific facts establishing that Defendant had pre-suit knowledge of the ’685 patent.

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

  • A core issue will be one of claim scope and construction: Does the claim term "subtracting a matrix from the kernel data" encompass the "refitting" process allegedly performed by NVIDIA's RAPIDS cuML, or does it require a more specific mathematical operation not present in the accused software?
  • A key evidentiary question will be one of technical implementation: What specific calculations does the accused RAPIDS cuML software perform when executing its Recursive Feature Elimination workflows, and do those calculations map onto the discrete steps recited in claim 1 of the ’685 patent?
  • A central legal question, potentially revisiting issues from prior litigation, will be patent eligibility under 35 U.S.C. § 101: Does claim 1 recite a patent-ineligible abstract mathematical algorithm, or is it directed to a specific improvement in the functionality of a computer-implemented learning system, as the complaint argues at length (Compl. ¶¶32, 38)?