Can the integration of Spatio-temporal continuity lead to more reliable and bio-plausible Capsule-based Networks when trained in an unsupervised way?

胶囊网络描述了一种最新的深度学习架构，其中单个捕获被单独训练，以鲁棒的方式提取特定的信息。然后，这些胶囊被整合到一个分层网络中，以解决更复杂的问题，这些问题太过具有挑战性，无法直接解决。这些网络显示出早期的希望，但尚未达到其理论性能水平。尼古拉的假设是，其中一个问题是，网络正在对独立的图像进行训练，这破坏了现实世界数据中固有的时空模式。这个博士项目将从人体皮质柱中获得灵感，为单个胶囊的设计提供信息，并在更现实的连续数据集上训练后评估它们的性能。

Capsule networks describe a recent deep-learning architecture in which individual captures are trained separately to extract a specific piece of information in a robust way. These capsules are then integrated in a hierarchical network to solve more complex problems that would be too challenging to solve directly. These networks show early promise but have yet to achieve their theoretical performance levels. Nicola's hypothesis is that one issue is that networks are training on independent images that destroy the spatio-temporal pattern inherent in real-world data. This PhD project shall take inspiration from the human cortical columns to inform the design of individual capsules and assess their performance when trained on more realistic continuous datasets.