Evolving Images for Visual Neurons Using a Deep Generative Network Reveals Coding Principles and Neuronal Preferences

What specific features should visual neurons encode, given the infinity of real-world images and the limited number of neurons available to represent them? We investigated neuronal selectivity in monkey inferotemporal cortex via the vast hypothesis space of a generative deep neural network, avoiding assumptions about features or semantic categories. A genetic algorithm searched this space for stimuli that maximized neuronal firing. This led to the evolution of rich synthetic images of objects with complex combinations of shapes, colors, and textures, sometimes resembling animals or familiar people, other times revealing novel patterns that did not map to any clear semantic category. These results expand our conception of the dictionary of features encoded in the cortex, and the approach can potentially reveal the internal representations of any system whose input can be captured by a generative model.

Two Types of Receptive Field Dynamics in Area V4 at the Time of Eye Movements?

How we perceive the world as stable despite the frequent disruptions of the retinal image caused by eye movements is one of the fundamental questions in sensory neuroscience. Seemingly convergent evidence points towards a mechanism which dynamically updates representations of visual space in anticipation of a movement (Wurtz, 2008). In particular, receptive fields (RFs) of neurons, predominantly within oculomotor and attention related brain structures (Duhamel et al., 1992; Walker et al., 1995; Umeno and Goldberg, 1997), are thought to "remap" to their future, post-movement location prior to an impending eye movement. New studies (Neupane et al., 2016a,b) report observations on RF dynamics at the time of eye movements of neurons in area V4. These dynamics are interpreted as being largely dominated by a remapping of RFs. Critically, these observations appear at odds with a previous study reporting a different type of RF dynamics within the same brain structure (Tolias et al., 2001), consisting of a shrinkage and shift of RFs towards the movement target. Importantly, RFs have been measured with different techniques in those studies. Here, we measured V4 RFs comparable to Neupane et al. (2016a,b) and observe a shrinkage and shift of RFs towards the movement target when analyzing the immediate stimulus response (Zirnsak et al., 2014). When analyzing the late stimulus response (Neupane et al., 2016a,b), we observe RF shifts resembling remapping. We discuss possible causes for these shifts and point out important issues which future studies on RF dynamics need to address.

Motion detection based on recurrent network dynamics

The detection of visual motion requires temporal delays to compare current with earlier visual input. Models of motion detection assume that these delays reside in separate classes of slow and fast thalamic cells, or slow and fast synaptic transmission. We used a data-driven modeling approach to generate a model that instead uses recurrent network dynamics with a single, fixed temporal integration window to implement the velocity computation. This model successfully reproduced the temporal response dynamics of a population of motion sensitive neurons in macaque middle temporal area (MT) and its constituent parts matched many of the properties found in the motion processing pathway (e.g., Gabor-like receptive fields (RFs), simple and complex cells, spatially asymmetric excitation and inhibition). Reverse correlation analysis revealed that a simplified network based on first and second order space-time correlations of the recurrent model behaved much like a feedforward motion energy (ME) model. The feedforward model, however, failed to capture the full speed tuning and direction selectivity properties based on higher than second order space-time correlations typically found in MT. These findings support the idea that recurrent network connectivity can create temporal delays to compute velocity. Moreover, the model explains why the motion detection system often behaves like a feedforward ME network, even though the anatomical evidence strongly suggests that this network should be dominated by recurrent feedback.

Cortical magnification plus cortical plasticity equals vision?

Most approaches to visual prostheses have focused on the retina, and for good reasons. The earlier that one introduces signals into the visual system, the more one can take advantage of its prodigious computational abilities. For methods that make use of microelectrodes to introduce electrical signals, however, the limited density and volume occupying nature of the electrodes place severe limits on the image resolution that can be provided to the brain. In this regard, non-retinal areas in general, and the primary visual cortex in particular, possess one large advantage: "magnification factor" (MF)-a value that represents the distance across a sheet of neurons that represents a given angle of the visual field. In the foveal representation of primate primary visual cortex, the MF is enormous-on the order of 15-20 mm/deg in monkeys and humans, whereas on the retina, the MF is limited by the optical design of the eye to around 0.3m m/deg. This means that, for an electrode array of a given density, a much higher-resolution image can be introduced into V1 than onto the retina (or any other visual structure). In addition to this tremendous advantage in resolution, visual cortex is plastic at many different levels ranging from a very local ability to learn to better detect electrical stimulation to higher levels of learning that permit human observers to adapt to radical changes to their visual inputs. We argue that the combination of the large magnification factor and the impressive ability of the cerebral cortex to learn to recognize arbitrary patterns, might outweigh the disadvantages of bypassing earlier processing stages and makes V1 a viable option for the restoration of vision.