Variational Inference on the Final-Layer Output of Neural Networks
This addresses the need for efficient uncertainty quantification in neural networks for applications requiring reliable predictions, though it is incremental by focusing on the output space rather than the full parameter space.
The paper tackles the problem of neural networks producing overconfident predictions by proposing Variational Inference on the Final-Layer Output (VIFO), which combines the simplicity of traditional training with Bayesian uncertainty quantification, resulting in improved performance and a good tradeoff in runtime and uncertainty, especially for out-of-distribution data.
Traditional neural networks are simple to train but they typically produce overconfident predictions. In contrast, Bayesian neural networks provide good uncertainty quantification but optimizing them is time consuming due to the large parameter space. This paper proposes to combine the advantages of both approaches by performing Variational Inference in the Final layer Output space (VIFO), because the output space is much smaller than the parameter space. We use neural networks to learn the mean and the variance of the probabilistic output. Using the Bayesian formulation we incorporate collapsed variational inference with VIFO which significantly improves the performance in practice. On the other hand, like standard, non-Bayesian models, VIFO enjoys simple training and one can use Rademacher complexity to provide risk bounds for the model. Experiments show that VIFO provides a good tradeoff in terms of run time and uncertainty quantification, especially for out of distribution data.