Date Published: July 25, 2013
Publisher: Public Library of Science
Author(s): Alex Roxin, Stefano Fusi, Jeff Beck
Abstract: Long-term memories are likely stored in the synaptic weights of neuronal networks in the brain. The storage capacity of such networks depends on the degree of plasticity of their synapses. Highly plastic synapses allow for strong memories, but these are quickly overwritten. On the other hand, less labile synapses result in long-lasting but weak memories. Here we show that the trade-off between memory strength and memory lifetime can be overcome by partitioning the memory system into multiple regions characterized by different levels of synaptic plasticity and transferring memory information from the more to less plastic region. The improvement in memory lifetime is proportional to the number of memory regions, and the initial memory strength can be orders of magnitude larger than in a non-partitioned memory system. This model provides a fundamental computational reason for memory consolidation processes at the systems level.
Partial Text: Memories are stored and retained through a series of complex, highly coupled processes that operate on different timescales. In particular, it is widely believed that after the initial encoding of a sensory-motor experience, a series of molecular, cellular, and system-level alterations lead to the stabilization of an initial memory representation (memory consolidation). Some of these alterations occur at the level of local synapses, while others involve the reorganization and consolidation of different types of memories in different brain areas. Studies of patient HM revealed that medial temporal lobe lesions severely impair the ability to consolidate new memories, whereas temporally remote memories remain intact . These results and more recent work (see e.g. ) suggest that there may be distinct memory systems, and that memories, or some of their components, are temporarily stored in the medial temporal lobe and then transferred to other areas of the cortex. Is there any fundamental computational reason for transferring memories from one area to another? Here we consider memory models consisting of several stages, with each stage representing a region of cortex characterized by a particular level of synaptic plasticity. Memories are continuously transferred from regions with more labile synapses to regions with reduced but longer-lasting synaptic modifications. Here we refer to each region as a stage in the memory transfer process. We find that such a multi-stage memory model significantly outperforms single-stage models, both in terms of the memory lifetimes and the strength of the stored memory. In particular, memory lifetimes are extended by a factor that is proportional to the number of memory stages.
In conclusion, we showed that there is a clear computational advantage in partitioning a memory system into distinct stages, and in transferring memories from fast to slow stages. Memory lifetimes are extended by a factor that is proportional to the number of stages, without sacrificing the amount of information stored per memory. For the same memory lifetimes, the initial memory strength can be orders of magnitude larger than in non-partitioned homogeneous memory systems. In the Results we focused on the differences between the heterogeneous and the memory system model. In Fig. S15 in Text S1 we show that the SNR of the memory transfer model (multistage model) is always larger than the SNR of homogeneous model for any learning rate. This is true also when one considers that homogeneous models can potentially store more information than the memory transfer model. Indeed, in the homogeneous model all synapses can be modified at the time of memory storage, not only the synapses of the first stage. However, the main limitation of homogeneous models with extended memory lifetimes comes from the tiny initial SNR. If one reduces the amount of information stored per memory to match the information stored in the memory transfer model, it is possible to extend an already long memory lifetime but the initial SNR reduces even further (see Text S1 for more details).
Here we include a brief description of the models and formulas used to generate the figures. For a detailed and comprehensive description of the models please refer to Text S1.