Closed-loop optimization of a high-dimensional generative latent space for rhythmic visual response

Neural oscillations accompany a wide range of cognitive states and behaviors including perception, memory, and movement, and modulating them is of growing interest for both basic neuroscience and clinical research. Previous demonstrations of closed-loop modulation of visual neural responses mainly relied on invasive recordings and focused on firing-rate maximization rather than rhythmic modulation. Here, we show that the relative power of steady-state visual evoked response (SSVEP), measured with readily-available, non-invasive electroencephalography, can be modulated in closed-loop as a function of image stimulus parameters for single participants within a single session. Stimulus optimization with flicker frequencies in the alpha and theta bands was successful in 10, 20, and 40 dimensional latent spaces. We also show that optimized stimuli generalize to new participants when shown in open-loop. Finally, we characterize the visual features that modulate relative SSVEP power and find that low-frequency spatial power in the image drives theta and alpha in opposite directions. Together, our results show that closed-loop stimulus optimization is a viable method for rhythmic neural modulation using noninvasive neuroimaging methods.