Optimal Stimuli as a New Method to Investigate Neural Networks - Abstract: Cortical microcircuits are cubic-millimeter neural networks that tile the brain surface and support every aspect of cortical function, from perception to motor skills to decision-making. As such, these circuits are theorized to form the fundamental “computational unit” in the brain, a flexible building block that can be repeated and recombined in new ways to achieve new functions. However, despite pioneering work investigating these microcircuits, we still have important gaps in knowledge about the computations they perform. Partially, these gaps come from limitations in traditional experimental designs. For example, in visual cortex, studies of cortical microcircuits have generally used single-modality stimuli, such as oriented gratings or colors. These simple stimuli only capture a neuron’s preferences for one type of visual feature and miss the diversity of visual features we observe in real life. In contrast, new machine learning approaches can gather data on a neuron and design “optimal” stimuli that incorporate a neuron’s preferences across all possible images. These approaches achieve significantly more precision than conventional methods and have the potential to finally unravel the inner workings of cortical microcircuits, leading to a more complete understanding of the brain for both science and medicine. However, to date, these new methods have only been applied to individual neural preferences: they have not been used to study how these neurons interact. Therefore, we will apply these new machine learning techniques to elucidate how the brain functions at a circuit level. Traditionally, columnar microcircuits are understood to have homogeneous stimulus preferences when using conventional stimuli. However, we know that these stimuli only capture about 45% of variability in the neurons’ responses: we do not know what the full range of natural images will reveal about these circuits. Therefore, due to differences in circuitry and cell types across cortical layers, we hypothesize that “optimal stimuli” will finally reveal heterogeneous stimulus preferences within cortical microcircuits, demonstrating the computations performed across these networks. Our technique combines an image generator, which produces a diverse range of color images, and an optimization algorithm which uses neural feedback to evolve these images and maximize the activity of a target neuron in visual cortex. Using this technique, we aim to characterize the precise computations performed by vertical connections across layers and horizontal connections across columns in the cortex. Overall, we propose an unprecedented study of cortical microcircuits using new machine learning techniques and offering both scientific and medical significance. Scientifically, our technique offers an invaluable tool for studying any stimulus-responsive brain region, providing a systematic way to characterize unknown brain circuits as well as reevaluating known brain circuits. Medically, this technique will empower the next generation of scientists to design more personalized interventions for psychiatric and neurological patients, uniquely targeted to their brain and neuropsychiatric profile.
