Aligning to learn

Various domains, including images, text, and brain activity, can be represented in embedding spaces. One question considered at this workshop is how to measure the alignment between such systems. For example, how aligned are human and model representations? However, this talk presents an alternative perspective by considering the potential of aligning systems to facilitate learning. We find that the natural environment furnishes the requisite information for aligning systems in an unsupervised manner. Children appear to align systems (e.g., text and images) when they learn labels for objects. Adults spontaneously align systems even when it's not helpful. One conclusion is that systems alignment complements standard event-based learning approaches.