The goal of computational neuroscience is to find mechanistic explanations of how the nervous system processes information to give rise to cognitive function and behavior. At the heart of the field are its models, that is, mathematical and computational descriptions of the system being studied, which map sensory stimuli to neural responses and/or neural to behavioral responses. These models range from simple to complex. Recently, deep neural networks (DNNs) have come to dominate several domains of artificial intelligence (AI). As the term “neural network” suggests, these models are inspired by biological brains. However, current DNNs neglect many details of biological neural networks. These simplifications contribute to their computational efficiency, enabling them to perform complex feats of intelligence, ranging from perceptual (e.g., visual object and auditory speech recognition) to cognitive tasks (e.g., machine translation), and on to motor control (e.g., playing computer games or controlling a robot arm). In addition to their ability to model complex intelligent behaviors, DNNs excel at predicting neural responses to novel sensory stimuli with accuracies well beyond any other currently available model type. DNNs can have millions of parameters, which are required to capture the domain knowledge needed for successful task performance. Contrary to the intuition that this renders them into impenetrable black boxes, the computational properties of the network units are the result of four directly manipulable elements: input statistics, network structure, functional objective, and learning algorithm. With full access to the activity and connectivity of all units, advanced visualization techniques, and analytic tools to map network representations to neural data, DNNs represent a powerful framework for building task-performing models and will drive substantial insights in computational neuroscience.
Tim C. Kietzmann, Patrick McClure, and Nikolaus Kriegeskorte
Tom Baden, Timm Schubert, Philipp Berens, and Thomas Euler
Visual processing begins in the retina—a thin, multilayered neuronal tissue lining the back of the vertebrate eye. The retina does not merely read out the constant stream of photons impinging on its dense array of photoreceptor cells. Instead it performs a first, extensive analysis of the visual scene, while constantly adapting its sensitivity range to the input statistics, such as the brightness or contrast distribution. The functional organization of the retina abides to several key organizational principles. These include overlapping and repeating instances of both divergence and convergence, constant and dynamic range-adjustments, and (perhaps most importantly) decomposition of image information into parallel channels. This is often referred to as “parallel processing.” To support this, the retina features a large diversity of neurons organized in functionally overlapping microcircuits that typically uniformly sample the retinal surface in a regular mosaic. Ultimately, each circuit drives spike trains in the retina’s output neurons, the retinal ganglion cells. Their axons form the optic nerve to convey multiple, distinctive, and often already heavily processed views of the world to higher visual centers in the brain. From an experimental point of view, the retina is a neuroscientist’s dream. While part of the central nervous system, the retina is largely self-contained, and depending on the species, it receives little feedback from downstream stages. This means that the tissue can be disconnected from the rest of the brain and studied in a dish for many hours without losing its functional integrity, all while retaining excellent experimental control over the exclusive natural network input: the visual stimulus. Once removed from the eyecup, the retina can be flattened, thus its neurons are easily accessed optically or using visually guided electrodes. Retinal tiling means that function studied at any one place can usually be considered representative for the entire tissue. At the same time, species-dependent specializations offer the opportunity to study circuits adapted to different visual tasks: for example, in case of our fovea, high-acuity vision. Taken together, today the retina is amongst the best understood complex neuronal tissues of the vertebrate brain.
Tatyana O. Sharpee
Sensory systems exist to provide an organism with information about the state of the environment that can be used to guide future actions and decisions. Remarkably, two conceptually simple yet general theorems from information theory can be used to evaluate the performance of any sensory system. One theorem states that there is a minimal amount of energy that an organism has to spend in order to capture a given amount of information about the environment. The second theorem states that the maximum rate with which the organism can acquire resources from the environment, relative to its competitors, is limited by the information this organism collects about the environment, also relative to its competitors. These two theorems provide a scaffold for formulating and testing general principles of sensory coding but leave unanswered many important practical questions of implementation in neural circuits. These implementation questions have guided thinking in entire subfields of sensory neuroscience, and include: What features in the sensory environment should be measured? Given that we make decisions on a variety of time scales, how should one solve trade-offs between making simpler measurements to guide minimal decisions vs. more elaborate sensory systems that have to overcome multiple delays between sensation and action. Once we agree on the types of features that are important to represent, how should they be represented? How should resources be allocated between different stages of processing, and where is the impact of noise most damaging? Finally, one should consider trade-offs between implementing a fixed strategy vs. an adaptive scheme that readjusts resources based on current needs. Where adaptation is considered, under what conditions does it become optimal to switch strategies? Research over the past 60 years has provided answers to almost all of these questions but primarily in early sensory systems. Joining these answers into a comprehensive framework is a challenge that will help us understand who we are and how we can make better use of limited natural resources.
John D. Medaglia and Danielle S. Bassett
Network analyses in nervous system disorders involve constructing and analyzing anatomical and functional brain networks from neuroimaging data to describe and predict the clinical syndromes that result from neuropathology. A network view of neurological disease and clinical syndromes facilitates accurate quantitative characterizations and mathematical models of complex nervous system disorders with relatively simple tools drawn from the field of graph theory. Networks are predominantly constructed from in vivo data acquired using physiological and neuroimaging techniques at the macroscale of nervous system organization. Studies support the emerging view that neuropsychiatric and neurological disorders result from pathological processes that disrupt the brain’s economically wired small-world organization. The lens of network science offers theoretical insight into progressive neurodegeneration, neuropsychological dysfunction, and potential anatomical targets for interventions ranging from pharmacological agents to brain stimulation.
Kenway Louie and Paul W. Glimcher
A core question in systems and computational neuroscience is how the brain represents information. Identifying principles of information coding in neural circuits is critical to understanding brain organization and function in sensory, motor, and cognitive neuroscience. This provides a conceptual bridge between the underlying biophysical mechanisms and the ultimate behavioral goals of the organism. Central to this framework is the question of computation: what are the relevant representations of input and output, and what algorithms govern the input-output transformation? Remarkably, evidence suggests that certain canonical computations exist across different circuits, brain regions, and species. Such computations are implemented by different biophysical and network mechanisms, indicating that the unifying target of conservation is the algorithmic form of information processing rather than the specific biological implementation. A prime candidate to serve as a canonical computation is divisive normalization, which scales the activity of a given neuron by the activity of a larger neuronal pool. This nonlinear transformation introduces an intrinsic contextual modulation into information coding, such that the selective response of a neuron to features of the input is scaled by other input characteristics. This contextual modulation allows the normalization model to capture a wide array of neural and behavioral phenomena not captured by simpler linear models of information processing. The generality and flexibility of the normalization model arises from the normalization pool, which allows different inputs to directly drive and suppress a given neuron, effectively separating information that drives excitation and contextual modulation. Originally proposed to describe responses in early visual cortex, normalization has been widely documented in different brain regions, hierarchical levels, and modalities of sensory processing; furthermore, recent work shows that the normalization extends to cognitive processes such as attention, multisensory integration, and decision making. This ubiquity reinforces the canonical nature of the normalization computation and highlights the importance of an algorithmic framework in linking biological mechanism and behavior.
Color perception in macaque monkeys and humans depends on the visually evoked activity in three cone photoreceptors and on neuronal post-processing of cone signals. Neuronal post-processing of cone signals occurs in two stages in the pathway from retina to the primary visual cortex. The first stage, in in P (midget) ganglion cells in the retina, is a single-opponent subtractive comparison of the cone signals. The single-opponent computation is then sent to neurons in the Parvocellular layers of the Lateral Geniculate Nucleus (LGN), the main visual nucleus of the thalamus. The second stage of processing of color-related signals is in the primary visual cortex, V1, where multiple comparisons of the single-opponent signals are made. The diversity of neuronal interactions in V1cortex causes the cortical color cells to be subdivided into classes of single-opponent cells and double-opponent cells. Double-opponent cells have visual properties that can be used to explain most of the phenomenology of color perception of surface colors; they respond best to color edges and spatial patterns of color. Single opponent cells, in retina, LGN, and V1, respond to color modulation over their receptive fields and respond best to color modulation over a large area in the visual field.
Anitha Pasupathy, Yasmine El-Shamayleh, and Dina V. Popovkina
Humans and other primates rely on vision. Our visual system endows us with the ability to perceive, recognize, and manipulate objects, to avoid obstacles and dangers, to choose foods appropriate for consumption, to read text, and to interpret facial expressions in social interactions. To support these visual functions, the primate brain captures a high-resolution image of the world in the retina and, through a series of intricate operations in the cerebral cortex, transforms this representation into a percept that reflects the physical characteristics of objects and surfaces in the environment. To construct a reliable and informative percept, the visual system discounts the influence of extraneous factors such as illumination, occlusions, and viewing conditions. This perceptual “invariance” can be thought of as the brain’s solution to an inverse inference problem in which the physical factors that gave rise to the retinal image are estimated. While the processes of perception and recognition seem fast and effortless, it is a challenging computational problem that involves a substantial proportion of the primate brain.
Luis Carrillo-Reid and Rafael Yuste
Despite over a century of neuroscience research, the nature of the neural code, that is, how neuronal activity underlies motor, sensory, and cognitive functions, remains elusive. Understanding the causal relation between neuronal activity and behavior requires a new conceptual paradigm that considers groups of neurons, instead of individual neurons, as the functional building blocks of the brain. These “neuronal ensembles,” defined as groups of neurons with coordinated activity that are reliably recalled by sensory stimuli, motor programs, or cognitive states, could be basic modular functional units of neural circuits. This hypothesis is consistent with past and present neuroscience results and could provide a broader framework to more effectively decipher the neural code in normal brains and provide new insights into how abnormal brain activity could lead to mental and neurological disease.