AI & Machine Learning

The hardest part is describing what the model actually does.

Machine learning applications sit on the sharpest edge of both eligibility and enablement. A claim to "a neural network that predicts X" is a claim to a result. What survives is a claim to the architecture, the training regime, or the systems engineering that made the result reachable.

Published patents in this area: 409
Professionals: 21

We work on model architectures and training methods, data pipeline and labeling infrastructure, inference and accelerator systems, computer vision, natural language and speech processing, and the applied machine learning inside medical and financial products. Where a trained model controls a plant, a robot or a vehicle, the model is claimed here and its use in Industrial Automation, Robotics or Autonomous Vehicles.

This area cuts across the firm's other technical groups rather than sitting beside them: an inference accelerator is a semiconductor matter, a training platform is a distributed systems matter, and a diagnostic model is a medical device matter. We staff it accordingly.

Claiming the method, not the result

A claim to a model that achieves an outcome is a claim to the outcome, and it will be met with both eligibility and enablement objections. What holds is the architecture, the training regime, the loss formulation, the data-handling step that made the outcome reachable, or the systems engineering that made it affordable at inference time.

This has a practical consequence for how applications are captured. The patentable contribution is frequently something the research team regards as an implementation detail rather than the result they are proud of, which means the intake conversation has to go past the paper abstract.

Enablement when the artifact is a trained model

Written description and enablement are strained by inventions whose behavior emerges from training rather than from design. A specification that describes only the architecture and asserts the result may not enable anything; the disclosure has to reach the training data characteristics, the hyperparameters, and the evaluation that makes the behavior reproducible — while remaining a document the client is willing to publish.

Balancing that against trade secret protection is a strategic decision, and it should be a deliberate one rather than a default.

Evidence

Patents we drafted in this area

Patents the firm prosecuted, drawn from its own published listing and from the USPTO's attorney-of-record index. A record, not a claim.

All patent listings

12,725,608Automatic speech recognition with multilingual scalability and low-resource adaptation
12,725,451Artificial intelligence video analysis enhanced with text generation
12,725,406Task-oriented clustering using prompt learning
12,725,076Artificial intelligence model learning introspection
12,725,075Training data generation via reinforcement learning fault-injection
12,725,036Explainable deep interpolation of missing pixels via non-contiguous interpolation neighborhoods
12,718,797Label smoothing technique for improving generalization of deep neural network acoustic models
12,718,138Quantum-enhanced features for classical machine learning