In this research meeting, we invited Gideon Kowadlo from Cerenaut.ai to talk about modelling of the hippocampus together with the neocortex for few-shot learning and beyond.
In mammalian brains, the neocortex and hippocampus are complementary modules that interact. Their interaction is known to be crucial in the formation of declarative memory as well as being important for Working Memory and executive control. Computational modelling of hippocampus and interaction between hippocampus and neocortex is of great importance to better understand neocortex itself, animal intelligence and to build more intelligent machines. A standard framework for hippocampal modelling is CLS. It captures an ability to learn distinct events rapidly. CLS has been tested on toy datasets, showing fast learning of specific examples, but not generalisation. In ML, the inverse is true. The standard approach to few-shot learning considers learning of categories, showing generalisation, but not instance learning (e.g. a particular tree), which is important for realistic agents. In addition, few-shot learning in ML is predominantly ‘short term’, without permanent incorporation of knowledge of new categories. We will describe extension to CLS, a novel Artificial Hippocampal Algorithm (AHA), which overcomes the above limitations.
“Unsupervised One-shot Learning of Both Specific Instances and Generalised Classes with a Hippocampal Architecture” paper: [2010.15999] Unsupervised One-shot Learning of Both Specific Instances and Generalised Classes with a Hippocampal Architecture
“One-shot learning for the long term: consolidation with an artificial hippocampal algorithm” paper: [2102.07503] One-shot learning for the long term: consolidation with an artificial hippocampal algorithm