1
views
0
recommends
+1 Recommend
0 collections
    0
    shares
      • Record: found
      • Abstract: not found
      • Article: not found

      Alleviating catastrophic forgetting using context-dependent gating and synaptic stabilization

      Read this article at

      ScienceOpenPublisherPMC
      Bookmark
          There is no author summary for this article yet. Authors can add summaries to their articles on ScienceOpen to make them more accessible to a non-specialist audience.

          Abstract

          <p id="d5647775e187">Artificial neural networks can suffer from catastrophic forgetting, in which learning a new task causes the network to forget how to perform previous tasks. While previous studies have proposed various methods that can alleviate forgetting over small numbers ( <span class="inline-formula"> <math id="i1" overflow="scroll"> <mo>⩽</mo> </math> </span>10) of tasks, it is uncertain whether they can prevent forgetting across larger numbers of tasks. In this study, we propose a neuroscience-inspired scheme, called “context-dependent gating,” in which mostly nonoverlapping sets of units are active for any one task. Importantly, context-dependent gating has a straightforward implementation, requires little extra computational overhead, and when combined with previous methods to stabilize connection weights, can allow networks to maintain high performance across large numbers of sequentially presented tasks. </p><p class="first" id="d5647775e197">Humans and most animals can learn new tasks without forgetting old ones. However, training artificial neural networks (ANNs) on new tasks typically causes them to forget previously learned tasks. This phenomenon is the result of “catastrophic forgetting,” in which training an ANN disrupts connection weights that were important for solving previous tasks, degrading task performance. Several recent studies have proposed methods to stabilize connection weights of ANNs that are deemed most important for solving a task, which helps alleviate catastrophic forgetting. Here, drawing inspiration from algorithms that are believed to be implemented in vivo, we propose a complementary method: adding a context-dependent gating signal, such that only sparse, mostly nonoverlapping patterns of units are active for any one task. This method is easy to implement, requires little computational overhead, and allows ANNs to maintain high performance across large numbers of sequentially presented tasks, particularly when combined with weight stabilization. We show that this method works for both feedforward and recurrent network architectures, trained using either supervised or reinforcement-based learning. This suggests that using multiple, complementary methods, akin to what is believed to occur in the brain, can be a highly effective strategy to support continual learning. </p>

          Related collections

          Most cited references 13

          • Record: found
          • Abstract: found
          • Article: not found

          Dynamic predictions: oscillations and synchrony in top-down processing.

          Classical theories of sensory processing view the brain as a passive, stimulus-driven device. By contrast, more recent approaches emphasize the constructive nature of perception, viewing it as an active and highly selective process. Indeed, there is ample evidence that the processing of stimuli is controlled by top-down influences that strongly shape the intrinsic dynamics of thalamocortical networks and constantly create predictions about forthcoming sensory events. We discuss recent experiments indicating that such predictions might be embodied in the temporal structure of both stimulus-evoked and ongoing activity, and that synchronous oscillations are particularly important in this process. Coherence among subthreshold membrane potential fluctuations could be exploited to express selective functional relationships during states of expectancy or attention, and these dynamic patterns could allow the grouping and selection of distributed neuronal responses for further processing.
            Bookmark
            • Record: found
            • Abstract: found
            • Article: found
            Is Open Access

            Dynamic reconfiguration of human brain networks during learning

             ,  ,   (2011)
            Human learning is a complex phenomenon requiring flexibility to adapt existing brain function and precision in selecting new neurophysiological activities to drive desired behavior. These two attributes -- flexibility and selection -- must operate over multiple temporal scales as performance of a skill changes from being slow and challenging to being fast and automatic. Such selective adaptability is naturally provided by modular structure, which plays a critical role in evolution, development, and optimal network function. Using functional connectivity measurements of brain activity acquired from initial training through mastery of a simple motor skill, we explore the role of modularity in human learning by identifying dynamic changes of modular organization spanning multiple temporal scales. Our results indicate that flexibility, which we measure by the allegiance of nodes to modules, in one experimental session predicts the relative amount of learning in a future session. We also develop a general statistical framework for the identification of modular architectures in evolving systems, which is broadly applicable to disciplines where network adaptability is crucial to the understanding of system performance.
              Bookmark
              • Record: found
              • Abstract: found
              • Article: not found

              Structure-stability-function relationships of dendritic spines.

              Dendritic spines, which receive most of the excitatory synaptic input in the cerebral cortex, are heterogeneous with regard to their structure, stability and function. Spines with large heads are stable, express large numbers of AMPA-type glutamate receptors, and contribute to strong synaptic connections. By contrast, spines with small heads are motile and unstable and contribute to weak or silent synaptic connections. Their structure-stability-function relationships suggest that large and small spines are "memory spines" and "learning spines", respectively. Given that turnover of glutamate receptors is rapid, spine structure and the underlying organization of the actin cytoskeleton are likely to be major determinants of fast synaptic transmission and, therefore, are likely to provide a physical basis for memory in cortical neuronal networks. Characterization of supramolecular complexes responsible for synaptic memory and learning is key to the understanding of brain function and disease.
                Bookmark

                Author and article information

                Journal
                Proceedings of the National Academy of Sciences
                Proc Natl Acad Sci USA
                Proceedings of the National Academy of Sciences
                0027-8424
                1091-6490
                October 12 2018
                : 201803839
                Article
                10.1073/pnas.1803839115
                6217392
                30315147
                © 2018

                Comments

                Comment on this article