Computational modeling of learning and adaptation in the visual system

Computational modeling of learning and adaptation in the visual system

Adaptation and life-long learning are hallmarks of human intelligent behavior, however, the neural mechanisms underlying different forms of adaptation and
experience-dependent modifications of cognitive capacities remain largely unknown. Although massive amount of neural data is generated by recent experimental
techniques, existing theories and models for the adaptability and development of cognitive functions are often too abstract to be directly compared to experimental data. In this project, we aim to use mathematical models and computer simulations, based on the known experimental data, to systematically investigate the adaptive learning behaviors of the visual system at three different levels and time scales. They are: 1) the developmental mechanism of the primary visual cortex of mice in the critical period induced by external stimuli; 2) the adaptive optimization of neural population coding strategy in the primary visual cortex of awake monkeys when they are trained to learn a two-week long perception task; and 3) the life-long adaptability of the face recognition system to cross-species faces in chimpanzees.

Understanding cognitive function of the brain is one of the most challenging tasks today. While a lot of research was done in Philosophy, Psychology, and Neuroscience over past centuries, the neuronal basis of many our cognitive abilities still remain a mystery. Humans interact with the environment in multiple ways, and to overcome challenges in every day’s life, our cognitive abilities need to be astounding. In every blink of a second, we effortlessly adapt to and perceive the visual scene, process this information online and in parallel to deduce the current situation, and combine this knowledge with our past experiences to finally generate adequate behavior.

Plastic changes of neural networks on multiple time scales are necessary and prerequisite for generating desired cognitive functions, and, in fact many neural systems in the brain are formed and structured in an experience-dependent manner (Holtmaat and Svoboda, 2009). Indeed, many neural systems have to remain plastic throughout life-time to be able to acquire knowledge or to adapt to long-term changes in the environments (Gould et al., 1999; Kempermann et al., 1997; Wiskott et al., 2006). Understanding the neural mechanisms of these adaptive processes is very important and may give insights to curing mental diseases (to retract “wrong” plastic changes) or disorders in the development of the cognitive abilities.

Today, higher cognitive functions are thought to arise from the complicated interplay between many dedicated brain areas (Van Essen et al., 1992), which themselves are vast networks of complex neurons having diverse morphologies and functions and communicating with electric and chemical signals of complex spatio-temporal dynamics (Buzsaki, 2006; Kandel et al., 2000). Not surprisingly, due to this complexity, neural mechanisms of many cognitive abilities to date remain unknown: integrating the vast amount of data necessary to understand the complex interplay of brain regions remains a big challenge.

That said, with the advent of very-high dimensional data recordings, such as multi-array neurophysiological recordings (Nordhausen et al., 1996), optical imaging techniques (Grinvald et al., 1986), fMRI (Logothetis et al., 1999), or population-wide two photon-calcium imaging (e.g.(Cossart et al., 2003)), today’s experimental techniques seem to be advanced enough to reach to an deeper understanding of the neural mechanisms of cognitive abilities.

However, without the help of modern computer models and simulations, together with new innovative data analysis tools, the experimentally observed neural data reflecting the complex processes leading to cognitive abilities are not likely to be deciphered.

Thus the hope and new approach of the neuroscience community is to combine experimentally obtained neural data, e.g. on the temporal dynamics of single neurons, into networks of neurons simulated on computers in order to ultimately understand cognitive functions (Gerstner et al., 2012). Recently, very large international research projects have been formed in this research area of computational modeling (Koch and Reid, 2012; Markram, 2006; Markram, 2012) (funded almost 1 billion euros in the next 10 years), underlining the importance and hope put into this approach by international researchers.

However, many experimental labs currently struggle to keep up with the demand of combining experimental and theoretical approaches, and as a result the combination of innovative analysis techniques and computational modeling very close to experimental data is still rare. In fact, this interdisciplinary computational approach is relatively young – especially so in China – because only recently computers and neuroscience techniques became powerful enough and because it requires both, mathematical and biological expertise, only achievable in close cooperation between experimentalists and modelers.

In this interdisciplinary project, we propose to form a unique strong union of modelers and experimental partners to develop computational models elucidating the neural correlates of a selection of adaptive processes. Our focus lies on the computational modeling of multiple types of experience-dependent network changes in and across the visual system, reaching from single neuron dynamics to behavioral levels.

Research on the visual system has a long history and the basic function of visual input processing is conceptually relatively well understood (Hubel and Wiesel, 1962; Kandel et al., 2000). Classically, the visual system is seen as (multiple) processing streams of hierarchically connected visual areas (Felleman and Van Essen, 1991; Grill-Spector and Malach, 2004; Mishkin et al., 1983; Van Essen et al., 1992), although newer research indicates that feed-back connections are also very important (Hupe et al., 2001; Rao et al., 1999; Sillito et al., 2006).

Although many details remain unclear (Olshausen and Field, 2005), the function of the early stages in the visual system is generally understood as a spatio-temporal filtering of the input stream, and neural activities are thought to reflect the localized luminance energy (Adelson and Bergen, 1985; Basole et al., 2003; Carandini et al., 1997; Mante and Carandini, 2005; Rasch et al., 2012). Moreover, diverse functions, such as contour integration (Field et al., 1993), border ownership signals (Zhou et al., 2000), contrast-gain control (Ohzawa et al., 1982), saliency maps (Li and others, 2002), or movement direction (Georgopoulos et al., 1986), are also attributed to the early visual cortices. Hierarchically higher visual areas commonly perform more and more abstract functions (Orban, 2008), ranging from motion extraction in MT (Newsome and Pare, 1988), and coding of corners and more complex shapes in V4 (Pasupathy and Connor, 1999), up to object detection in IT (Kobatake and Tanaka, 1994), and dedicated areas for face feature extraction, such as the fusiform face area (Kanwisher et al., 1997; Tsao et al., 2003).

While the general function of many visual areas is approximately understood, development and experience-dependent plasticity of the visual cortex is less well studied and neural correlates remain unclear. It is known that shortly after birth in mice, the visual system needs to mature in an activity-dependent manner (Huberman et al., 2008; Levelt and Hubener, 2012; Movshon and Van Sluyters, 1981). Additionally, even in the adult visual cortex, adaptation and experience-dependent changes are various and happen on multiple time scales, ranging from milliseconds, e.g. the frequency adaptation of single neurons to constant inputs (Connors and Gutnick, 1990), to minutes, e.g. the synaptic strength adaption called short-term plasticity (Zucker and Regehr, 2002) or contrast adaptation (Carandini and Ferster, 1997), up to hours and days, e.g. perceptual learning in V1 (Fahle, 2002; Li et al., 2004; Sasaki et al., 2009), and years, e.g. the other-species effect we recently found in face perception or chimpanzees (Dahl et al., 2013).

Adaptation and long-lasting plastic, experience-dependent changes – all present in the visual system – may have very different neural mechanisms. Today, although often experimental details as well as principled theoretical models exist, there remains a large gap in understanding between these conceptual, very abstract theoretical models of experience-dependent adaptation and neural activity observed in experiments. The proposed project plans to develop computational models and analysis tools of experimental data at hand to bridge this gap to reach a better understanding of the neural mechanisms of adaptive cognitive abilities and adaptation in the visual system.

The project will comprise computational modeling of three types of experience-dependent changes in the visual cortex having diverse time scale and scope; these are: A) the drastic developmental network modifications during maturation of the primary visual cortex, B) the modest changes of the neural activity during perceptual learning of the early visual cortex, and C) the life-long adaption of cognitive abilities and higher visual functions as exemplified in the adaptation of the face processing system.

These selections of experience-dependent changes in the visual cortex are chosen because: 1) while conceptual ideas exist in the literature, the neural mechanism remain largely unknown; 2) the prior expertise of our experimental cooperation partners and the availability of sufficient neural data to guide a computational modeling process close to experimental observations; and 3) these three topics present a spectrum of adaptation processes in the visual system ranging from very early developmental effects, later modifications in and across the visual hierarchy (perceptual learning), to adaption of higher-order cognition and thus will provide a bigger picture of neural mechanisms of adaptations in the visual cortex.