Neuro AI has made strides in machine intelligence by harnessing insights from biological intelligence for building better machines. Inspired by the normative principles underlying the nonlinear response properties of neurons in the visual cortex, we take a prominent class of deep generative models (Variational Autoencoders, or VAE for short) and apply the neuroscience insight to a long-standing challenge in the VAE literature: improving its knowledge on what it does not know. Overconfidence is not a trait we usually cherish, and in machine intelligence tools such overconfidence can be dangerous for error-critical applications. It is this overconfidence that is characteristic to VAEs: when learning to interpret a specific set of images, VAEs have the important ability to assign a confidence to the interpretation. However, when presenting them with an image that is way beyond the expertise, VAEs tend to misinterpret the image ad doing so with an unimpressive confidence.
In a collaboration with the Echeveste lab (sinc(i), CONICET-UNL, Santa Fe, Argentina) we propose that a twenty-something year old idea, the infinite scale mixtures can fix erroneous inference. As a bonus, we show that a widespread phenomenon in the biological brain, divisive normalization naturally emerges in this artificial system and contributes to addressing the challenge.
The study was spearheaded by Jose Catoni and Domonkos Martos. The paper is appearing in IEEE Transaction on Pattern Analysis and Machine Intelligence, preprint is available at arXiv:
Remedying uncertainty representations in visual inference through Explaining-Away Variational Autoencoders
Josefina Catoni, Domonkos Martos, Ferenc Csikor, Enzo Ferrante, Diego H. Milone, Balázs Meszéna, Gergő Orbán, Rodrigo Echeveste
arXiv:2404.15390, https://doi.org/10.48550/arXiv.2404.15390.
Caption for the cover image: the feature giving off broken inference is the lack of ability of learning a bow-tie pattern in the latent space (the space designed to interpret observations) characteristic of natural images. The infinite scale mixtures trick is specifically addressing this bow-tie pattern in Variational Autoencoders.