UC Berkeley

Visual cognition applies temporal smoothing to its inputs, which creates a serial dependence between successive representations at the focus of attention. This is thought to promote perceptual stability. While the benefits of serial dependence have been assumed, evidence that perception itself is altered has been limited. In the first chapter of this dissertation, I vary the delay between stimulus and response in a spatial delayed response task to investigate whether serial dependence occurs at the time of perception or later in working memory. I find that behavioral responses made immediately after viewing a stimulus are on average veridical. Only as memory demands increase is a blending of past and present information apparent in behavior, reaching its maximum with a memory delay of six seconds.

In the second chapter, I explore potential neural-circuit mechanisms of serial dependence. I consider two possible substrates of the effect: stable persistent activity during the memory delay and dynamic “activity-silent” synaptic plasticity. I find that networks endowed with both strong reverberation to support persistent activity and dynamic synapses can closely reproduce behavioral serial dependence. Specifically, elevated activity drives synaptic augmentation, which biases activity on the upcoming trial, giving rise to a spatiotemporally tuned shift in the population response. My hybrid neural model is a theoretical advance beyond abstract mathematical characterizations of working memory and demonstrates the power of biological insights to provide a quantitative explanation of human behavior.

The model developed in Chapter 2 proposes that serial dependence is due in part to synaptic augmentation, which is especially prominent in prefrontal cortex. In the third chapter, I investigate whether the bias in behavior depends on activity in three separate nodes of prefrontal cortex (PFC) – the frontal eye fields, the dorsolateral PFC, and the anterior PFC near the frontal pole. I find that transcranial magnetic stimulation (TMS) to these nodes causes reductions in serial dependence consistent with the model’s predictions. In contrast, TMS to posterior sites – either primary somatosensory cortex or posterior parietal cortex – fails to alter the magnitude of the behavioral effect. This general result holds across TMS protocols (online vs. offline) and tasks with different stimulus and response types.