Automated AI for Adapting to New Data Without Retraining

This patent describes an automated system that builds artificial intelligence models capable of adapting to new, different data without needing full retraining, by learning to ignore irrelevant changes.

What it covers

This system automatically builds artificial neural network architectures. It uses interfaces to receive training, validation, and testing data, which includes 'task labels Y' (what the AI needs to identify) and 'nuisance variations S' (irrelevant changes, like lighting or background). Memory banks store 'reconfigurable deep neural network (DNN) blocks' and settings called 'hyperparameters'. A processor then explores different hyperparameters and methods, including 'auxiliary regularization modules', to adjust how the DNN blocks work. The goal is to make the AI's predictions 'insensitive to the nuisance variations S', meaning it can identify the task labels accurately even when the irrelevant parts of the data change. For example, an AI trained to recognize a specific object in bright light could still recognize it in dim light without needing to be fully retrained.

What it doesn't cover

• —Does not cover manually designed neural network architectures for transfer learning, as it specifies 'automated construction'.
• —Does not cover transfer learning methods that do not explicitly disentangle latent variables from nuisance variations using auxiliary regularization modules.
• —Does not cover systems that do not explore hyperparameters of regularization, pre-processing, or post-processing methods to achieve nuisance robustness.
• —Does not cover non-deep neural network (DNN) architectures, as it specifically references 'reconfigurable deep neural network (DNN) blocks'.
• —Does not cover methods that do not aim for 'nuisance-robust Bayesian inference' to handle domain shifts.

The clever bit

The innovation lies in automatically exploring and configuring neural network components and methods to actively 'disentangle' important features from irrelevant 'nuisance variations'. This allows the AI to learn what truly matters for a task, making it highly adaptable and robust when encountering new, slightly different data environments.

Why it matters

AI models often struggle when deployed in real-world situations where data differs slightly from their training data, a problem known as 'domain shift'. This patent offers a way to automatically build AI models that are robust to these changes, making them more reliable and reducing the need for expensive, time-consuming retraining. This is crucial for deploying AI in diverse environments, from self-driving cars encountering varied weather to medical imaging systems used in different hospitals.

Real-world examples

• 1.Autonomous vehicle perception systems adapting to different weather or lighting conditions
• 2.Medical image analysis systems working across various scanner types or patient demographics
• 3.Industrial inspection systems identifying defects on different production lines
• 4.Speech recognition systems handling varying background noise or speaker accents