How an AI Learns to Recognize Objects by Generating Its Own Examples

This patent describes an enhanced artificial intelligence system that uses two competing networks, a generator and a discriminator, with special 'capsule layers' to create and recognize specific types of data, like images, even with limited initial examples.

What it covers

This patent details an 'enhanced generative adversarial network' (GAN) and a method for recognizing specific 'target samples.' The system includes at least one 'enhanced generator' that creates new data from initial data, and at least one 'enhanced discriminator' that evaluates this generated data. The discriminator then provides feedback to the generator, helping it improve. A key feature of this enhanced discriminator, as described in claim 1, is its structure: a 'convolution layer,' a 'basic capsule layer,' a 'convolution capsule layer,' and a 'classification capsule layer,' all connected in sequence. The number and setup of these layers can be adjusted based on the specific features of the 'target sample' being recognized. For example, to recognize a rare type of medical image, the system could generate many synthetic versions of that image, and the discriminator, with its capsule layers, would learn to distinguish them from other images, even if only a few real examples exist.

What it doesn't cover

• —Does not cover generative adversarial networks that do not include capsule layers within their discriminator component.
• —Does not cover target sample recognition methods that do not use an enhanced generative adversarial network with the specified discriminator architecture.
• —Does not cover GANs where the discriminator's convolution, basic capsule, convolution capsule, and classification capsule layers are not sequentially connected.
• —Does not cover systems where the quantities and structure parameters of the discriminator's specific layers are not set according to the target sample's features.

The clever bit

The truly clever part is the integration of capsule networks directly into the discriminator of a GAN. While GANs are good at generating and distinguishing data, capsule networks excel at understanding hierarchical relationships and orientations of features, allowing the discriminator to learn more robust and invariant representations of 'target samples' than traditional convolutional neural networks.

Why it matters

Generative Adversarial Networks (GANs) are a powerful class of artificial intelligence models used for generating realistic data, like images, audio, or text. This patent enhances the discriminator part of a GAN by incorporating 'capsule layers,' which are known for better understanding spatial relationships and orientations in data, especially images. This improvement could lead to more robust and accurate recognition systems, particularly in fields where data is scarce or complex, such as medical diagnostics or specialized object detection.

Real-world examples

• 1.Advanced medical image analysis for disease detection
• 2.High-fidelity synthetic data generation for training AI models
• 3.Facial recognition systems that need to identify faces from various angles
• 4.Quality control systems for manufacturing that detect subtle defects
• 5.Autonomous vehicle perception systems for recognizing specific objects