The dynamics of category-specific perception in ventral temporal cortex
Miller, Kai Joshua
MetadataShow full item record
My thesis investigates the relationship between the ventral temporal cortex and perception of different categories of visual stimuli, in humans. Under the guidance of my scientific and clinical mentors, we performed a set of visual perception experiments with human patients who had been implanted with arrays of subdural electrocorticography (ECoG) electrodes on the ventral temporal brain surface. In order to understand these experiments and what they teach us about brain function, my evolving thought process throughout graduate school was guided by a series of central questions: What are the relevant portions of the ECoG signal needed to understand ventral temporal physiology? The text begins with a general discussion of category-selective regions in the ventral temporal brain and introduces some of the principles of ECoG measurement. The ECoG signal reflects multiple aspects of population-scale neurophysiology, and these are carefully explored and illustrated specifically in the ventral visual stream. "Broadband" changes in the ECoG voltage power spectrum are decoupled from intermingled oscillatory motifs, and shown to be a generic correlate of local cortical function across a variety of brain areas and behavioral tasks. Continuous approximations of this broadband track behavior at approximately the 20-50ms timescale, making it the optimal setting to explore visual processing, perception, and decision making on the same timescales that these phenomena occur. Likewise, the rhythmic gamma oscillation has been proposed to be a key marker of visual perception and attention. We find that robust gamma oscillations can indeed be generated with specific visual stimuli, such as luminance gratings, but are not a prerequisite for perception, as they are not seen with noise patterns, nor with most easily identifiable natural images. However, many brain rhythms are more than just passive phenomena in visual cortex. By examining how broadband spectral change is entrained on the phase low frequency oscillations (δ, θ, α, β, and γ), we were able to show that brain rhythms selectively couple to the average population activity in a task-dependent manner. This entrainment mechanism dynamically engages and releases cortical areas on a second-to-second basis during visual processing. Because this rhythmic influence in the α and β ranges is selectively decreased in early visual areas during visual task engagement, we propose a "suppression by synchronization" hypothesis, where widespread populations of cortical neurons are synchronized and phase-coupled to the rhythm to impose a cortically suppressed (disengaged) state. These rhythmic changes, like the stimulus-triggered averaged potentials (ERPs), are idiosyncratic to specific cortical loci, and not generic across brain regions. Broadband spectral changes, in contrast, are a generic correlate of local activity in each cortical region. How robust are the neuronal population responses that we measure? In order to quantify how reliable the ECoG response is to individual types of stimuli in ventral temporal cortex, we showed subjects simple grayscale pictures of faces and houses. Robust category-specific responses to each type of stimulus were found, where face-picture selective sites were consistently found on the fusiform gyrus and house-selective sites were found on the lingual and parahippocampal gyri. A novel method was developed, where averaged ECoG response templates (both broadband and ERP) from a training period were projected into a testing period. Our initial decoding allowed for pre-designated stimulus timing, and 97% of stimuli could be accurately classified as face or house. Real world experience, however, is spontaneous, and to this end, we attempted to simultaneously predict the occurrence, timing, and type of visual stimuli from the spontaneous ECoG stream. In this setting, 96% of all stimuli were captured spontaneously, with ~20ms timing error (only 4% of the spontaneous predictions were incorrect; 4% of the stimuli were missed). This examination of the spatiotemporal distributions of category-selective cortical sites reveals that activity highly delineated and can be classified robustly, and activity temporally is sufficient to construct a simple visual stream from the ECoG array, at the timescales of perception. Do these ventral temporal neuronal populations reflect perceptual context? Even in a series of randomly interleaved grayscale faces and houses, there is perceptual context. A face following a prior face is viewed in a different context than a face that follows a house, and we isolated face- and house- selective ventral temporal loci to see if this sequence-context might be reflected in the single-trial broadband responses. Approximately half of the category-selective sites showed greater total broadband activity for novel than for repeated stimulus class. In approximately half of the face-selective sites (and none of the house-selective sites), latency to peak activity was faster for novel than repeat stimuli (with no correlation between the total activity and latency effects). These observations suggest that aspects of ventral temporal cortical physiology are especially tuned to sequential context, with an emphasis on novelty during perception. The unexpected finding that novel stimuli are processed faster suggests that ventral temporal cortex is optimized for perceiving changing context. Is ventral temporal cortex is an active locus of visual object perception? A face and house picture decision task was performed, where visual evidence was parametrically degraded, and responses were measured from category-selective loci. Robust neurometric functions revealed decreasing neuronal population activity as image noise increased, with shape preserved whether subjects correctly reported the stimulus or not. A basic face-house template-projection classifier was built on a simple, choice-free, localizer experiment, and then applied to the decision task where visual evidence was parametrically degraded. At low levels of stimulus noise, predicted stimulus type was correlated more closely with the stimulus than subjects' perception. The converse was true at high noise level, where predicted stimulus type was more correlated with the subjects' choice than stimulus type. With decreasing stimulus (increasing noise) the influence of higher order cognition is revealed. Ventral temporal cortex must therefore be an active locus of perception, where bottom-up and top-down information converge.