MIT researchers have developed a unique technique that enables a machine-learning model to quantify its confidence in its predictions. This technique requires a manageable amount of new data and is significantly less computationally intensive than other techniques.

“No one has a clue how to build a conscious machine, at all.” - Stuart Russell

Overview

Machine-learning models help people solve complex problems, like finding diseases in medical pictures or obstacles on the road so those self-driving cars can avoid them. But machine-learning models can be wrong, so people must know when to believe a model's predictions in high-stakes situations.

Uncertainty quantification is one way to make a model more reliable. When the model makes a prediction, it gives a score showing how confident the forecast is correct. Unfortunately, even though figuring out how much uncertainty there is can be helpful, most current methods involve restraining the whole model from giving it that ability. For example, training a model means showing it millions of examples of how to do a job to learn it. Retraining requires millions of new data inputs and a lot of computing power.

Uncertainty learning

When using the output label distribution directly to generate uncertainty measures, neural networks are known to be overconfident. Existing techniques mainly address this issue by retraining the entire model to impose uncertainty quantification capacity, allowing the learnt model to attain targeted accuracy and uncertainty prediction performance simultaneously. However, training the model from scratch is time-consuming and may need to be more practical. 

The researchers suggest a unique Bayesian metamodel that is both effective and computationally economical for augmenting pre-trained models with superior uncertainty quantification abilities. Furthermore, their recommended method requires no additional training data. It is versatile enough to assess various uncertainties and quickly adapt to multiple application contexts, such as out-of-domain data detection, misclassification detection, and trustworthy transfer learning. 

Quantifying uncertainty

In uncertainty quantification, a machine-learning model assigns a numeric score to each output that reflects the model's confidence in the accuracy of that prediction. However, incorporating uncertainty quantification by building a new model or retraining an existing model requires a lot of data and expensive computation, making it impractical. Moreover, existing methods sometimes have the unintended effect of diminishing the accuracy of model predictions.

Data uncertainty is produced by corrupted data or incorrect labelling, and it can only be minimised by repairing the dataset or collecting new data. Model uncertainty occurs when the model is unsure how to explain freshly observed data and may generate inaccurate predictions, most frequently due to a lack of similar training instances. It is a complicated yet common issue when models are deployed. They often encounter data that differs from the training sample in real-world scenarios.

Validation

Once a model generates an uncertainty quantification score, the user must still have confidence in the score's accuracy. Typically, researchers validate the accuracy of a model by creating a subset of the original training data and evaluating the model on this subset. They developed a new validation technique by adding noise to the validation set's data; this chaotic data is more comparable to out-of-sample data that can cause model uncertainty. This chaotic dataset is used to evaluate uncertainty quantifications by the researchers.

Conclusion

The researchers show that their suggested metamodel approach is flexible and outperforms numerous representative image classification benchmarks in these applications. Furthermore, this technique could assist researchers in enabling more machine-learning models to perform uncertainty quantification effectively, ultimately assisting users in determining when to trust predictions.

Image source: Unsplash