Verified by 
Artificial Intelligence
New AI Learns to Evaluate Its Own Decisions
MIT and Harvard’s deep neural network rapidly gauges confidence in AI decisions.
Posted December 29, 2020 Reviewed by Kaja Perina
Earlier this month, researchers at the Massachusetts Institute of Technology (MIT) and Harvard University presented a novel method for machine learning to gauge the confidence of its own predictions at the 34th Conference on Neural Information Processing Systems (NeurIPS). The researchers presented a faster, more efficient way for regression-based neural networks (NNs) to reduce uncertainty, which could help improve machine learning safety and performance.
Artificial intelligence (AI) deep learning and the biological brain share a fundamental problem—the complex internal processing of data is largely a mystery to scientists, akin to the proverbial black-box. Moreover, machine learning is prone to fragility, as its quality is impacted by data volume, bias, and relevance, as well as the creation and tuning of the AI algorithms.
“Uncertainty estimation for neural networks has very significant societal impact,” wrote the research team of Alexander Amini, Daniela Rus, and Wilko Schwarting at the MIT Computer Science and Artificial Intelligence Lab (CSAIL), along with Ava Soleimany at the Harvard Biophysics Graduate Program. “Neural networks are increasingly being trained as black-box predictors and being placed in larger decision systems where errors in their predictions can pose immediate threat to downstream tasks.”
The researchers’ approach was to create a method that could estimate uncertainty from just one run of the neural network that produces a decision, along with a probabilistic distribution called an evidential distribution because it has the evidence that supports the decision.
To achieve this, they trained non-Bayesian neural networks to estimate a continuous target and the associated evidence to learn the uncertainty of the data and statistics (aleatoric uncertainty), as well as the uncertainty of the model itself (epistemic uncertainty). Then the researchers placed evidential priors over the original Gaussian likelihood function, versus placing priors on network weights which is done for Bayesian neural networks. A Gaussian distribution is also called a normal distribution, or bell curve.
The neural network is trained to infer and output the hyperparameters of the higher-order evidential distribution. Hyperparameters are parameters that are set prior to machine learning, with values that cannot be estimated from the data, and are external to the model.
This enables the representation of the uncertainty to be learned without requiring sampling during inference or on out-of-distribution examples for training. Priors are imposed during training to regularize the model whenever the predicted evidence is unaligned with the desired output. Overall, the method is efficient because it only needs a single forward pass through the deterministic neural network to estimate uncertainty.
To test their method, the researchers trained a neural network on a high-dimensional computer vision task to estimate depth values for monocular color images for each pixel. The training data used included more than 27,000 color image pairs of indoor scenes such as rooms in a house.
“Evidential models exhibit strong performance, as error steadily decreases with increasing confidence,” observed the researchers.
Next, the team performed out-of-distribution testing by inputting outdoor driving images from the ApolloScape dataset. This type of testing helps the researchers find out how well the model performs with out-of-distribution data. Up until now, the model has only seen indoor images that it was trained on prior. The evidential model showed increased uncertainty on the outdoor driving images.
Finally, the researchers input images that were slightly altered to observe how the model would perform in extreme cases of out-of-distribution data with deliberately perturbed images. The evidential model showed the ability to identify and reflect greater predictive uncertainty on the perturbed images.
“Not only does the predictive uncertainty steadily increase with increasing noise, but the spatial concentrations of uncertainty throughout the image also maintain tight correspondence with the error,” the researchers observed.
Currently, artificial intelligence machine learning is being deployed across industries for a myriad of purposes such as financial services, synthetic biology, quantum chemistry, RNA-based biotechnology tool, battery designs for electric cars, early warning for extreme weather, antibiotics drug discovery, and even for early detection of diseases as a linguistic biomarker. AI computer vision is an especially useful tool for robotics, health care diagnostics, and autonomous vehicles, where a high degree of accuracy matters.
“The efficiency, scalability, and calibration of our approach could enable the precise and fast uncertainty estimation required for robust NN deployment in safety-critical prediction domains,” wrote the researchers.
This new technique enables deep neural networks to provide confidence levels in its decisions in a rapid, more efficient way, to shed some light on the opacity of the black-box of artificial intelligence.
Copyright © 2020 Cami Rosso All rights reserved.
