Introduction

Synaptic Plasticity

The brain is a computational marvel, performing millions of operations per second that allow you to speak, read this text, ride a bike, and so much more. The brain's computing units, cells called neurons, signal to one another at specialized structures called synapses (Figure 1). Synapses are critical for coordinating the multi-neuron activity necessary in complex information processing operations, such as motion or sensation. [1]

Figure 1: A schematic of a chemical synapse, the most common type of synapse in the human brain. The pre-synaptic neuron (top) undergoes an action potential which leads to neurotransmitter release into the synaptic cleft. Neurotransmitters bind corresponding receptors on the post-synaptic neuron (bottom) which can lead to various signaling consequences. Figure adapted from Pereda (2014). [2]

However, if synapses were static, unchanging structures then it would be nearly impossible for the brain to adapt to new experiences due to fixed "connection weights" between different neurons. Thankfully, synapses are dynamic structures that can undergo several processes collectively termed synaptic plasticity to alter neural connectivity based on experience. [3]

The most well-understood of these synaptic plasticity concepts is the idea of Hebbian learning which was first introduced as a simple theory in 1949: neurons that fire together, wire together. [4, 5] Bliss and Lomo (1973) proved that this phenomenon occurs in physiological neurons through a process called long-term potentiation, or LTP (Figure 2). [6] Mechanisms of synaptic plasticity, such as LTP, have been implicated in various neural processes. However, plasticity is usually explored in relation to learning and memory due to the theoretical importance of "retuning" neural weights in these processes. [3]

Figure 2: An example time series of fEPSP slope (representative of post-synaptic activity) in an LTP experiment. At 10 minutes, high-frequency stimulation gives rise to a prolonged increase in fEPSP slope. Example fEPSPs 1 and 2 are shown at 10 minutes and 50 minutes, respectively. Figure adapted from Citri and Malenka (2008). [3]

Learning, Memory, and Cascade Models

Learning and memory are topics that have been explored extensively in the neuroscience literature due to their importance in a wide variety of everyday tasks such as object detection, speech, and navigation. Although the processes of learning and memory seem functionally similar, Fusi et al. (2005) explain that the processes of learning and memory theoretically require different degrees of plasticity. To rapidly learn new information from experience, the brain should be incredibly plastic in order to encode novel relationships. Conversely, to robustly retain memories of prior experiences, the brain should be incredibly rigid in order to prevent new information from "overwriting" old memories. However, cognitive psychology studies have shown the capacity for humans to display both strong initial learning as well as the robust storage of prior memories. [7]

Fusi et al. (2005) demonstrate how traditional theoretical models of memory that view synaptic plasticity as a binary switch between a "weak" and "strong" synapse cannot simultaneously result in strong learning and robust memory storage due to the exponential forgetting curves generated by these models. Instead, the authors follow prior work in cognitive psychology which had shown human forgetting curves to be well-fit by a power-law relationship. [8] Based on generating a power-law forgetting curve, they develop the cascade model of synaptic plasticity, where two synaptic states, weak or strong, are characterized by further substates that can be achieved via metaplasticity mechanisms (Figure 3). The authors show that this model generates a power-law forgetting curve, while also showing a capacity for simultaneous strong initial learning and robust memory storage that had not been displayed by prior models. [7]

Figure 3: Schematic of cascade plasticity with two synaptic states: weak (brown) and strong (teal). Each synaptic state is characterized by 5 substates. Vertical transitions (substate to substate) are indicative of metaplasticity within a synaptic state and occur with probabilities p. Horizontal transitions (synaptic state to synaptic state) are indicative of normal plasticity and occur with probabilities q. Figure taken from Fusi et al. (2005). [7]

Research Questions

  1. Can memory storage lifetimes be decreased by latent processes, such as teaching inputs, that increase the probabilities of plasticity events occurring?

  2. How do asymmetric cascades between synaptic states affect memory storage lifetimes?