Final Activity Report Summary - PLICON (Plasticity and Learning in Cortical Networks)
The basis of the brain's remarkable learning abilities are thought to be so-called plasticity mechanisms which change the properties of individual nerve cells and of the synapses connecting them. Many such mechanisms have been discovered and studied, but their relation to learning phenomena at the level of an organism's behaviour is still largely unclear and extremely difficult to study. Previous research has typically assumed that individual plasticity mechanisms can be directly related to specific learning abilities. But an alternative view is that the brain's learning abilities arise from the interaction of several plasticity mechanisms.
The goal of the PLICON project was to study through theoretical analysis and computer simulations, how the interaction of different forms of plasticity may give rise to powerful learning abilities. We investigated this issue in the context of several kinds of artificial neural networks: so-called recurrent networks that are suited for processing time-varying input signals and feed forward networks that learn to encode static input patterns. In both cases we could show how powerful learning abilities arise from the interaction of multiple plasticity mechanisms and how each individual plasticity form was critical for the overall behaviour.
Specifically, we showed the critical importance of homeostatic plasticity mechanisms that regulate the amount of input to a neuron, its amount of firing, or even the amount of information transmitted by a neuron. Another line of research connected the cellular and synaptic mechanisms of learning to the behavioural level. Specifically, we investigated through what learning mechanisms human infants may develop an understanding of the world around them.
To this end, we developed neural network models that learned early concepts, such as object unity, the three-dimensional nature of space, object occlusions, and causality, as well as simple forms of social interactions. In several of these models, reward-driven learning played a central role, highlighting the importance of brain-internal value systems for our cognitive development. A second theme of central importance turned out to be the ability to make predictions about the world. Overall, we showed that several fundamental concepts or building blocks in our "cognitive vocabulary" can result from generic learning mechanisms.
Our results represent an important existence proof that human infants may learn such concepts with existing plasticity mechanisms and that these concepts do not have to be innate, as has been claimed in the past. A final line of research studied the neural basis of short-term or working memory, our ability to remember stimuli for brief periods of time. We performed psychological experiments studying the effects of holding visual objects in memory on classifying other objects during the retention period. In parallel, we studied how working memory may form during development. To this end we studied how recurrent networks using multiple forms of plasticity and reward-dependent learning can develop the ability to store stimuli for short periods of time.
The goal of the PLICON project was to study through theoretical analysis and computer simulations, how the interaction of different forms of plasticity may give rise to powerful learning abilities. We investigated this issue in the context of several kinds of artificial neural networks: so-called recurrent networks that are suited for processing time-varying input signals and feed forward networks that learn to encode static input patterns. In both cases we could show how powerful learning abilities arise from the interaction of multiple plasticity mechanisms and how each individual plasticity form was critical for the overall behaviour.
Specifically, we showed the critical importance of homeostatic plasticity mechanisms that regulate the amount of input to a neuron, its amount of firing, or even the amount of information transmitted by a neuron. Another line of research connected the cellular and synaptic mechanisms of learning to the behavioural level. Specifically, we investigated through what learning mechanisms human infants may develop an understanding of the world around them.
To this end, we developed neural network models that learned early concepts, such as object unity, the three-dimensional nature of space, object occlusions, and causality, as well as simple forms of social interactions. In several of these models, reward-driven learning played a central role, highlighting the importance of brain-internal value systems for our cognitive development. A second theme of central importance turned out to be the ability to make predictions about the world. Overall, we showed that several fundamental concepts or building blocks in our "cognitive vocabulary" can result from generic learning mechanisms.
Our results represent an important existence proof that human infants may learn such concepts with existing plasticity mechanisms and that these concepts do not have to be innate, as has been claimed in the past. A final line of research studied the neural basis of short-term or working memory, our ability to remember stimuli for brief periods of time. We performed psychological experiments studying the effects of holding visual objects in memory on classifying other objects during the retention period. In parallel, we studied how working memory may form during development. To this end we studied how recurrent networks using multiple forms of plasticity and reward-dependent learning can develop the ability to store stimuli for short periods of time.