Improving AI Predictions by Factoring in How Much You Trust Each Model

This patent describes a method for making an AI system more accurate by combining several individual AI models and giving more weight to the ones considered more trustworthy.

What it covers

This patent describes a method for training an ensemble machine learning model, which is like a team of AI models working together. First, it receives data about how much to trust each individual AI model in the team (Claim 1a). This 'trust score' can be based on things like how sensitive the model is to bad input data or how confident its past predictions were (Claim 5). Next, it calculates an estimated prediction error for each model, but this error is adjusted by its trust score (Claim 1b). A less trusted model might have its errors penalized more. Then, it uses these adjusted error estimates to figure out a 'normalized weight' for each model (Claim 1c). Finally, it uses these normalized weights to determine the overall prediction equation for the entire ensemble, giving more influence to the more trusted models (Claim 1d). For example, if you have three AI models predicting stock prices, and one is known to be very reliable with certain types of data, this method would give that reliable model's predictions more say in the final combined forecast.

What it doesn't cover

• —Does not cover single machine learning models that do not combine predictions from multiple individual models.
• —Does not cover ensemble models that combine predictions without explicitly calculating and using a 'trust score' for each individual model.
• —Does not cover methods where the trust score does not influence the prediction error estimate or the normalized weights of the individual models.
• —Does not cover trust scores that are not real numbers, specifically those outside the 0.0 to 1.0 range as described in Claim 3.
• —Does not cover ensemble methods where the output prediction equation is not determined, at least in part, by these normalized weights.

The clever bit

The clever part is explicitly integrating a 'trust factor' into the mathematical calculation of how much each individual model's prediction contributes to the final combined output. Instead of just relying on past performance metrics, this method allows for a more nuanced understanding of a model's reliability to directly influence its weighting.

Why it matters

As AI models become more complex and are used in critical applications like healthcare or finance, ensuring their reliability and accuracy is crucial. This patent offers a way to build more robust AI systems by explicitly incorporating a measure of trust into how multiple models collaborate. This can lead to more dependable predictions, especially when individual models might have varying levels of quality or applicability to different data.

Real-world examples

• 1.Fraud detection systems combining multiple AI models
• 2.Medical diagnostic tools using various AI algorithms
• 3.Autonomous vehicle perception systems blending sensor data interpretations
• 4.Financial forecasting platforms integrating diverse predictive models
• 5.Cybersecurity threat detection systems