Memory retention the synaptic stability versus plasticity dilemma Paper: Abraham, Wickliffe C., and Anthony Robins. "Memory retention the synaptic stability versus plasticity dilemma." Trends in neurosciences 28.2 (2005): 73-78. 27.3.2017 Jerguš Greššák
Introduction? What role does synaptic stability and synaptic plasticity play in memory retention in neuronal circuits. Regulated balance of synaptic stability and synaptic plasticity is required for optimal memory retention.
Synaptic configuration In Hebb s theory, psychological event is pattern of neuronal activity within cell assemblies. To remember a previously experienced event, its activity pattern needs to be re-established. Activity pattern can be for example represented as specific setting of synaptic connection weights (synaptic configuration). Simple model: Each memory is encoded with specific synaptic configuration Long-term memory (LTM) equate with the persistence of a synaptic configuration.
Synaptic basis of memory retention Stability hypothesis: Strong version: One-to-one correspondence between a memory activity pattern and a fixed synaptic weight configuration in the circuits that encode it. Weaker version: Allow for some weakening and strengthening of weights over time. When synaptic weights change memory retrieval is impaired. Plasticity hypothesis: Weights encoding a memory can change. Memory is retained as long as the functionality (the ability to recreate the activity pattern) is preserved.
Support for the synaptic stability hypothesis? Can LTP in hippocampus last long enough to support very long-term memory (vltm). LTP in the hippocampus can reliably persist for days or weeks. After that in most studies LTP declines to baseline. It is consistent with theories that the hippocampus has a time-limited role in memory consolidation. Recent study: LTP in the rat hippocampal dentate gyrus have the capacity to last stably for many months or longer.
Induced LTP rat experiment Induction decremental and stable LTP in the dentate gyrus of awake freely moving rats. HFS high-frequency stimulation EE - enriched environment containing novel objects, a novel food and conspecific animals
Spine morphology mice experiment Spine presence stability support for synaptic stability hypothesis Two-photon laserscanning imaging in transgenic mice. Spines imaged across a 1 month interval 96% of spines in visual cortex were retained and overall persisted with a half-life > 13 months
Support for the synaptic plasticity hypothesis Memory consolidation theory: information moves from a temporary holding store to an anatomically separate permanent store. Experimental reactivation of previously consolidated memory makes it labile to disruption and makes it possible to be reconsolidated. Possible plasticity updates during sleep or related behaviours.
NMDA receptor and synaptic plasticity NMDA receptor in dynamically maintaining the long-term synaptic stability of memory storage circuits in the brain. Experiments on transgenic mice: 1. Contextual fear and water-maze training 2. Researchers switched off NMDA receptor in the forebrain during the storage stage 3. Impaired memory in mice Consolidation of memory in the early post-training period and the stable maintenance of remote memory require continued neural activity and NMDA receptor activation.? Ongoing plasticity merely refreshes the weights for the synapses or there are more widespread changes in synaptic structure and function.
Plasticity in spine morphology experiment Time-lapse imaging: Somatosensory cortex: 17% of spines had lifetimes 1 day 23% had a mean lifetime of 2 3 days Visual cortex: 96% of spines had estimated half-lives of 13 months changes in length or diameter of a significant proportion of these spines over a one-month interval? Plasticity stability dilemma still remains an experimentally unresolved issue for neurobiologists.
Artificial neural networks simulation Backpropagation learning algorithm Standard learning algorithm Training information (set of activity patterns) is presented as whole. Information is processed in one session until a correct output pattern is given for each input. Training is finished and network remains frozen. Sequential learning algorithm New information can be integrated with old information at any time. Incorporate partial rehearsal of old patterns while learning new ones. Anthony Robins described a effective sweep rehearsal algorithm.
Backpropagation network IH - input hidden layer connections HO - hidden output layer connections Network used in simulation: 32 input units 16 hidden units 32 output units
Simulation Standard backpropagation vs. backpropagation with rehearsal Base population of 10 activity patterns Activity pattern - randomly generated real valued inputs and associated outputs 40 new patterns were trained one at a time. Trained using either standard backpropagation methods or backpropagation with sweep rehearsal of old patterns.
Simulation Stability of connection weights Standard algorithm: Rapid loss of memory despite high correlation Rehearsal algorithm: Network must compute a new configuration each time a new item is learned
ANN simulations summary Standard ANN relative weight stability is associated with memory loss in the face of new learning. Preserving the contents of memory during ongoing learning requires an active maintenance process (via rehearsal) and flexibility in the connection weights. They predict that specific memory patterns in the brain can be encoded by different synaptic weight configurations. That is important to accommodate new learning.
Synthesis of the synaptic stability and plasticity hypotheses Synaptic plasticity Memory storage in the mammalian brain is generally thought to involve synaptic plasticity in extensive neural networks, cortical and subcortical. To preserve old information in the face of new learning, memory-storing synapses must retain the capacity for plasticity. Synaptic stability Stability in the weights throughout the memory network is needed during the intervals between learning or rehearsal episodes.
Support of existing theories Hebb s theory: Activity pattern within the cell assembly as the defining event for memory expression, not the specific synaptic configuration. Conceivable that there are multiple solutions for the pattern of synaptic weights that will reinstate the key activity pattern within a cell assembly. Confirmed in the ANN modelling. Other theories state: The need for the neocortex to be a slow learner so that information can be incorporated into the existing structure without causing loss of earlier learned information.
Thank you for your attention