Deep neural networks (DNNs) are often framed as being inspired by the brain. Although their advent has led to remarkable success stories and a plethora of large-scale applications, they however lack a key ability humans posses: the capacity for lifelong learning. That is, DNNs are primarily successful when trained on predefined training and dedicated test datasets. When faced with the breath of novelty and changing experiences in the real world, they become prone to erratic predictions and suffer from catastrophic forgetting. In contrast to the brain’s remarkable adaptation capabilities, this leads to a current practice of frequent and unsustainable re-training. In this talk, I will introduce the elements necessary to shift such prevalent static design towards lifelong machine learning systems; systems that transcend stationary datasets and continually learn in a world full of unknowns. I will outline mechanisms that augment DNNs with the ability to dynamically adapt their structure, robustly deal with novel situations, and efficiently incorporate new knowledge over time. Finally, I will conclude on why these newfound capabilities are but a few of the exciting avenues towards lifelong neural networks, drawing brief parallels to the host of mechanisms posited to contribute to lifelong learning in the brain.

Lecturer:
Prof. Dr. Martin Mundt

Contact:
Agnes Janßen (ajanssen@neuro.uni-bremen.de)