Patent Landscape:
Machine Learning

Neural Network Learning by Back Propagation

Trustees of Boston University

One of the earliest patents on backpropagation — the algorithm that made multi-layer neural networks trainable at scale. Backpropagation is the foundational mechanism behind every deep learning model trained today, from image classifiers to large language models.

Convolutional Neural Network for Image and Video Analysis

Google

Google's CNN architecture patent covers the specific layer configuration and weight-sharing methods used in its image classification systems. Convolutional networks are the backbone of computer vision — this patent's claims define how filters are applied across spatial dimensions to extract hierarchical features.

Generative Adversarial Network Training Method

NVIDIA

NVIDIA's GAN training patent covers the adversarial optimization process where a generator and discriminator network compete to produce synthetic data indistinguishable from real samples. GANs power deepfake detection, image synthesis, drug discovery, and data augmentation — making this one of the most commercially significant ML patents filed in the 2010s.

Recurrent Neural Network with Attention Mechanism

Google Brain

Covers the attention mechanism applied to recurrent architectures — the precursor insight to the full transformer model. Attention allows models to selectively weight different parts of an input sequence, solving the bottleneck problem that limited earlier sequence-to-sequence models in translation and summarization tasks.

Federated Machine Learning with Privacy-Preserving Aggregation

Google

Federated learning allows ML models to be trained across distributed devices without raw data ever leaving those devices. Google developed this approach for training Gboard on users' phones. This patent covers the aggregation protocol that combines model updates while preserving differential privacy — now standard in healthcare and finance ML applications.

Sparse Mixture-of-Experts Neural Network Architecture

Google Brain

The Mixture-of-Experts (MoE) architecture activates only a subset of model parameters for each input — enabling dramatically larger models without proportional compute costs. MoE is the architecture underlying GPT-4 and Gemini Ultra, making this patent directly relevant to the frontier model race.

Google / DeepMind

Holds the deepest ML architecture patent portfolio across reinforcement learning, transformer variants, and large-scale training methods. Google deliberately left the original transformer paper (Vaswani et al., 2017) unpatented — but filed aggressively on surrounding applications. DeepMind's AlphaFold protein structure prediction patents represent a separate IP moat in scientific ML.

NVIDIA

The hardware-software integration of CUDA with ML frameworks is NVIDIA's primary IP moat. Beyond GPU architecture patents, NVIDIA holds significant IP in training optimization methods, inference acceleration, and distributed training across GPU clusters. The move into DGX cloud infrastructure is protected by an expanding systems patent portfolio.

IBM Research

IBM's ML patent portfolio exceeds 9,000 filings with emphasis on enterprise applications: explainable AI, bias detection, and automated machine learning (AutoML). IBM's strategy is to license into regulated industries — healthcare, finance, legal — where explainability requirements create demand for its specific IP.

Microsoft

Uses OpenAI's open publication strategy as a complement — OpenAI publishes research, Microsoft patents the commercial implementations. Azure ML infrastructure, GitHub Copilot's code completion pipeline, and Office 365 AI integration are all covered by distinct Microsoft ML patent filings.