Autonomous Learning of Object Representations

• Applicant: Professor Dr. Jochen Triesch
• Subject Area: Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing ; General, Cognitive and Mathematical Psychology ; Experimental and Theoretical Network Neuroscience
• Term: since 2022
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 459426179

One of the first and most important feats of abstraction of our brains is to recognize objects independent of perspective, distance, background, and lighting. This achievement relies on information processing in the so-called ventral stream of our brain’s visual system, which is home to a hierarchy of representations of different levels of abstraction. This processing stream has been well studied, yet it is still unclear how this and similar feats of abstraction are learned. Today’s artificial object recognition systems based on deep neural networks exhibit a similar hierarchical structure, but they learn very differently. In contrast to their biological model, they often require millions of labeled training examples, while a few, or sometimes even a single example, suffice for us. How is this possible and can we mimic this much more efficient learning in artificial systems? This project brings together classic ideas from Neuroscience and modern machine learning approaches and proposes a new architecture, the so-called “contrastive learning through time” (CLTT), to bridge the gap between biological and artificial vision systems. At the same time, it develops a new approach for active learning of object representations driven by curiosity to allow future artificial vision systems to learn much more autonomously. Together with our collaborators from the ARENA consortium we will directly compare the learning and the representations resulting from CLTT with that of human subjects in psychological experiments. Overall, the project will help us better understand how our brain learns abstract representations so efficiently and how we can endow technical systems with the same ability.

• DFG Programme: Research Units
• Subproject of: FOR 5368: KI-FOR Abstract Representations in Neural Architectures (ARENA)