Models that impact people’s lives must be honest and forthcoming about their doubts. This new method brings balance to explainable AI.
At Featurespace, we have published a method for explaining model uncertainties in deep networks. Our approach makes it possible to understand the confounding patterns in the input data that a model believes do not fit with the class it predicts. This allows us to deploy honest and transparent models that give balanced and complete evidence for a decision, rather than just curating the evidence in support of its decision.
For example, in the figure below we observe three images of celebrities contained within the CelebA dataset. In all cases, the celebrities are labelled as “not smiling”, which is easily predicted using widely available deep classification models for image processing tasks, such as VGG16, ResNet50 or EfficientNet. However, models can struggle to offer confident classifications for some of these pictures. In the examples below, our proposed method highlights the various features that caused models to be uncertain (right column). We readily notice the presence of smile arcs, upper lip curvatures and buccal corridors commonly associated with smiles and grins.
Uncertainty and Bayesian Deep Learning
We built our uncertainty attribution method on top of Bayesian neural networks, (BNNs). This framework thoroughly treats all of the sources of uncertainty in a prediction including the uncertainty emanating from the model choice and limitations of the training data, which emerges as uncertainty on the fitted weight parameters of every neural cell. A good overview of BNNs is available here. In summary:
- Every fitted parameter for a neuron is captured as a probability distribution that represents our lack of certainty regarding its best value, given the training data.
- Since a deep model has plenty of parameters, thus, we have a multivariate distribution over all trained parameters, referred to as the posterior distribution.
- When we classify a new data point, each plausible combination of fitted parameter values offers us a different score! Rather than a single score, the BNN offers us a distribution of possible scores, called the posterior predictive distribution.
- BNNs commonly return the mean of the…
Continue reading: https://towardsdatascience.com/how-to-make-artificial-intelligence-articulate-doubt-9c2885a9e541?source=rss—-7f60cf5620c9—4