Humans quickly parse and categorize stimuli by combining perceptual information and previously learned knowledge. We are capable of learning new information quickly with only a few observations, and sometimes even a single observation. This one-shot learning (OSL) capability is still very difficult to realize in machine learning models. Novelty is commonly thought to be the primary driver for OSL. However, neuroscience literature shows that biological OSL mechanisms are guided by uncertainty, rather than novelty, motivating us to explore this idea for machine learning.
In this work, we investigate OSL for neural networks using more robust compositional knowledge representations and a biologically inspired uncertainty mechanism to modulate the rate of learning. We introduce several new neural network models that combine Holographic Reduced Representation (HRR) and Variational Autoencoders. Extending these new models culminates in the Holographic Generative Memory (HGMEM) model.
HGMEM is a novel unsupervised memory augmented neural network. It offers solutions to many of the practical drawbacks associated with HRRs while also providing storage, recall, and generation of latent compositional knowledge representations. Uncertainty is measured as a native part of HGMEM operation by applying trained probabilistic dropout to fully-connected layers. During training, the learning rate is modulated using these uncertainty measurements in a manner inspired by our motivating neuroscience mechanism for OSL. Model performance is demonstrated on several image datasets with experiments that reflect our theoretical approach.
Computer Engineering (MS)
Department, Program, or Center
Computer Engineering (KGCOE)
Graham, Dillon R., "Holographic Generative Memory: Neurally Inspired One-Shot Learning with Memory Augmented Neural Networks" (2018). Thesis. Rochester Institute of Technology. Accessed from
RIT – Main Campus