Shane McGlaun

Deep learning neural networks are artificial intelligence systems that are being used for increasingly important decisions. Deep learning neural networks are used for tasks as varied as autonomous driving to diagnosing medical conditions. This type of network excels at recognizing patterns in large and complex datasets to help with decision-making.

One big challenge is determining if the neural network is correct. Researchers at MIT and Harvard University have developed a quick way for a neural network to churn through data and provide a prediction along with the neural network’s confidence level in its answer. Researchers on the project believe that their system could save lives since deep learning is already deployed in the real world.

Currently, uncertainty estimation for neural networks tends to be computationally expensive and too slow for split-second decisions. The approach devised by the researchers is called “deep evidential regression” and speeds the process up, potentially leading to safer outcomes. Researchers on the project say we need the ability to have high-performance models and understand when results from the models can’t be trusted.

Deep learning has demonstrated impressive performance for a variety of tasks. In some instances, it has been able to surpass human accuracy. These networks are good at knowing the right answer 99 percent of the time, but any error is unacceptable with lives on the line. The researchers have devised a way to estimate uncertainty using a single run of the neural network. The network was designed with an increased output producing a decision and a new probabilistic distribution capturing evidence in support of its decision.ADVERTISING

The distributions are called evidential distributions and directly capture the model’s confidence in its prediction. Researchers tested their system using a challenging computer vision task. They trained the neural network to analyze a binocular color image and estimated the distance value for each pixel. This is a task that an autonomous vehicle might perform. The new network’s performance was on par with previous state-of-the-art models and added the ability to estimate its uncertainty.