No More AI Shortcuts in Future, More Reliable Forecasts
With the new approach, the machine learning (ML) prototype targets more data while task learning results in more reliable forecasts. In ML, while making a decision, the prototype depends on a simple dataset characteristic - shortcut solution occurs, which results in incorrect predictions.
More Data in Decision-making
MIT researchers introduce a new study by exploring the issue of ML shortcuts and put forward a solution that averts shortcuts by forcing the prototype to focus on more complex aspects, leveraging more data for decision-making. This will improve the model's efficiency in detecting diseases in medical images. However, shortcuts in such a situation might have dangerous implications for patients.
For a massive range of data, self-supervised ML is used, where the models are trained leveraging raw data that doesn't have people's label descriptions. But if the model uses shortcuts and malfunctions to get essential data, then these tasks can't use that data either. Another robust form of self-supervised ML - contrastive learning leverages an encoder algorithm to differentiate pairs of similar and dissimilar inputs.
The encoder must target valuable data characteristics for decision-making, but in this case, the method has fallen victim to shortcut solutions. While the work continues taking some crucial approaches towards understanding shortcuts, researchers said that the refinement of solutions and application to other kinds of self-supervised ML will be the next step to the future.