Introducing MA-PaPSP.
A Visual-oriented Hallucination Detector.
Project Highlights
- Selective prediction is a classic problem in machine learning.
- The objective is to achieve high accuracy on the predictions it does make.
- The objective is to achieve high accuracy on the predictions it does make.
Motivation for Research.
Selective prediction is a classic problem in machine learning in which a model is allowed to decide when to make a prediction and when to abstain. The objective is to achieve high accuracy on the predictions it does make, while avoiding uncertain or potentially risky inputs. Traditionally, one common (and widely used) approach relies on thresholding the model’s output confidence, often derived from logits or predicted probabilities. For instance, consider a classifier that must distinguish between a dog and a cat image. Now suppose the model produces a probability distribution over the two classes of say, 0.1 for dog and 0.9 for cat. This maximum probability (0.9) reflects the model’s confidence in its prediction. In selective prediction, we compare this confidence score to a predefined threshold (e.g., 0.6) and output the predicted class (cat here) when the confidence exceeds the threshold. Otherwise, the model simply abstains from making a prediction, signaling uncertainty.
Keep Reading.
Traditionally, one common (and widely used) approach relies on thresholding the model’s output confidence, often derived from logits or predicted probabilities.
Selective prediction is a classic problem in machine learning in which a model is allowed to decide when to make a prediction and when to abstain. The objective is to achieve high accuracy on the predictions it does make, while avoiding uncertain or potentially risky inputs.
Selective prediction is a classic problem in machine learning in which a model is allowed to decide when to make a prediction and when to abstain.
The objective is to achieve high accuracy on the predictions it does make, while avoiding uncertain or potentially risky inputs.