Deep Learning

Deep learning is the branch of machine learning that uses neural networks with many layers. GAO describes its rise this way: "Deep learning systems, which consist of neural networks that contain a large number of hidden layers, became prevalent in the 2010s and led to advancements in computer vision and natural language processing" (GAO-24-106946, 2024).

Neural networks

GAO says neural networks "are modeled loosely on the human brain and recognize patterns in data," and that they "consist of an arrangement of interconnected nodes."

Common network types

• Feedforward networks. NIST defines these as "Artificial neural networks in which the connections between nodes is from one layer to the next and do not form a cycle" (NIST AI 100-2e2025 glossary).
• Convolutional neural networks (CNNs). In convolutional layers, "feature detectors (known as kernels or filters) detect specific features across the input data." NIST says CNNs "are primarily used for processing grid-like data, such as images, and are particularly effective for tasks like image classification, object detection, and image segmentation." See Computer Vision.
• Transformers. GAO calls these "a type of neural network that applies widely to natural language processing by tracking the relationships between words within sentences to learn context and meaning." Transformers underlie most large language models; see Large Language Model.

Weaknesses NIST documents

• Adversarial examples. NIST AI 100-2e2025 recounts research that "showed that deep neural networks used for image classification could be easily manipulated through adversarial examples," meaning inputs altered slightly to change the output.
• Scale. NIST AI 100-1 notes that "many systems contain billions or even trillions of decision points," which makes behavior hard to trace.
• Unpredictable failures. NIST lists a "Higher degree of difficulty in predicting failure modes for emergent properties of large-scale pre-trained models."
• Opacity. General AI risks, including reproducibility, are covered under Artificial Intelligence.

Why it matters for land and mining claim records

Scanned plats, recorded pages, and historical maps are images, the grid-like data NIST says CNNs are built for. That is why deep learning appears in document and map tools. A deep network does not give a line-by-line reason for its reading, so the check on its output is the source image itself: what the model read from a page should be compared against the page.