The Simons Collaboration on the Global Brain
Traditionally, neuroscience research has relied heavily on studies of sensory and motor systems to reveal the principles of brain function. But much of what goes on in the brain is internal — states that control motivation and decision bias, representations of remembered events and cognitive explorations. These internal states play against and interact with sensory and motor signals in a way that has been difficult to explore in the absence of methods for reading out and affecting neural responses that do not rely solely on manipulating stimuli or monitoring actions. The time has come to extend our domain of study to include internal activity because the right methods now exist: We can record simultaneously from large numbers of neurons, we can manipulate their activity with optogenetics and we have the theoretical tools to decipher the signals that they encode. This allows us to recognize internal brain states from activity alone, unprecedented in the history of brain research. By combining this analysis of internal states with the analysis of sensory and motor processing, we can exploit new experimental and theoretical tools to span the arc from sensation to action, and to discover the nature, role and mechanisms of the neural activity that produces cognition.
1. Neural Coding and Dynamics in the Collaboration on the Global Brain
The term ‘global brain’ refers to the large-scale activity of neural populations within and across different regions of the brain. The complexity and richness of global brain activity has been strikingly visualized in recent large-scale optical recordings at cellular resolution from populations of neurons in the awake brain. Patterns of activity fluctuate across populations of cells in local neural circuits, while global flashes across multiple areas represent other forms of internal dynamics and brain states of unknown significance. These phenomena are but an initial glimpse of what we can expect to see in the near future using large-scale recording from behaving animals, ranging from flies to rodents to nonhuman primates. The scientific question addressed by the Simons Collaboration on the Global Brain (SCGB) is: How does this complexity and richness of brain activity represent and produce cognition?
A foundation for characterizing and understanding global brain activity is provided by the concepts of neural coding and neural dynamics. ‘Neural coding’ refers to how information (sensory input, emotional state, plans for future actions, motor output, etc.) is represented in the electrical and chemical activity of the neurons in a circuit. It is increasingly recognized that most neural codes must be understood at the multi-neuron level as patterns of activity across a population of cells. ‘Neural dynamics’ refers to how these coded representations change with time to produce cognition and behavior — the patterns of activity across neural populations as these evolve over time in the process of perception, thinking, and action.
The SCGB focuses on revolutionizing our understanding of neural coding and dynamics by exploiting new recording technologies and complementary mathematical analyses, and using this information to ultimately produce a mechanistic understanding of brain function. Large-scale recording using the new technologies will be combined with statistical analysis to investigate how neural coding and dynamics represent and process information relevant to behavior. For example, what is the form of neural dynamics in a circuit that makes a decision, and how is the information relevant to making that decision expressed in the neural code? What are the dynamically changing patterns of activity for recalling a memory, imagining the future, or, at the human level, performing mental arithmetic? The neural coding and dynamics we characterize will provide the basis for theory and modeling: The goal is to determine how the biophysical properties of neurons and the architecture of circuit connections determine these codes and dynamics, thus providing a mechanistic understanding of brain function.
The SCGB seeks general principles. In the same way that the dynamic electrophysiological properties of neurons (action potentials, interactions of excitation and inhibition, bursting and adaptation, etc.) generalize across brain areas and species, it is likely that fundamental forms of neural coding dynamics will also generalize. A primary goal of the collaboration is to identify and characterize these universal forms, because they represent general principles of brain function.
Revealing the general principles of coding and dynamics in neural circuits will not only provide insight into the neural basis of simple behaviors and cognitive phenomena, but it will ultimately revolutionize our understanding of everything from individual differences in behavior — why does a mathematician excel at abstract reasoning while a tennis player has superb motor skills? — to really understanding what is different or defective in the brains of individuals suffering from mental illness.
A paradigm shift for neural coding and dynamics
A technological revolution is occurring in our ability to monitor and manipulate brain activity. Particularly exciting advances are occurring at the “mesoscopic scale” of neural circuits and neural populations. This scale is intermediate between single cells, traditionally studied with single-unit microelectrodes, and entire brain regions, where activity has typically been mapped and studied using indirect or volume-averaged methods such as fMRI, MEG and EEG. The goal of the new technologies is maintaining the cellular resolution of single-unit studies but dramatically increasing the number of units simultaneously monitored so that entire local circuits or whole brain areas can be studied at once. Advances in photonics and imaging in conjunction with new generations of genetically encoded sensors of calcium and voltage now allow hundreds to thousands of neurons to be monitored simultaneously at cellular and single-action-potential resolution. New virtual-reality systems and miniaturized head-mounted microscopes provide ways for applying these optical measurements of population activity in the awake brains of rodents and other species performing behaviors such as decision making and navigation. In the small brains of worms, flies and other species, monitoring all of the brain’s neurons at once is now feasible. In parallel, dense microelectrode recording using microfabricated silicon probes and new multiple-probe arrays has become a practical reality; these are particularly useful in studies of neural coding and dynamics in nonhuman primates and humans. This technological revolution in monitoring neural populations is disruptive in the sense that it is facilitating a true paradigm shift in attacking the problem of neural coding and dynamics and the mechanisms of brain function.
The paradigm shift is occurring because much brain activity is not well captured in current studies. For one thing, elements of sensory, cognitive and motor function may be difficult or impossible to discern from the small-scale recordings that have been the mainstay of electrophysiology. This is because information is typically encoded across neural populations and can only be interpreted from analysis of the multi-neuron patterns of activity. In addition, the spontaneous, free-flowing nature of mental experience is not well captured when technical limitations require time-locked averaging of signals across trials, sessions and subjects, which is performed to reduce noise in the face of spike variability in order to make sense of single-unit recordings. The new population imaging and electrode-recording techniques, combined with new methods for analysis of data, promise to liberate us from this methodological straitjacket. Brain states and neural coding of information that can only be deduced from analysis of a population of cells become accessible, both for post hoc analysis of information coding and dynamics and for use in real-time applications. Real-time applications include neural prostheses and closed-loop perturbation experiments where, for example, sensory stimuli or optogenetic stimulations are triggered by the detection of a particular brain state.
Complementing the new technologies for large-scale recording are equally exciting new methods, such as optogenetics, for stimulating neurons in order to test the causal role of observed coding and dynamics. In the near future, we anticipate the ability not only to activate specific neural populations with light but to do so in specific identified cell-by-cell spatial patterns and sequences — in other words, neural activation at cellular resolution, to complement neural recording at cellular resolution. When fully developed, this technology should allow the experimenter to elicit particular brain states and test their causal role in cognition.
2. Extracting Knowledge and Understanding from Brain Activity
The experiments being proposed in the Simons Collaboration on the Global Brain will generate large quantities of data from extremely complex systems. The goals of the collaboration cannot be achieved without developing and applying methods of analysis to extract basic mechanisms and underlying principles from these data. Analysis in this context means both transforming the data into forms more suitable for study and using mathematical and computer-based models to test various explanatory hypotheses. Data transformations consist of extracting probability distributions, signals of relevance and dimensionally reduced descriptions from the data. Modeling consists of constructing networks that account for the data on the basis of biophysically realistic mechanisms and providing predictions for further experimental study.
Probabilistic descriptions, signal processing and data reduction are classic methods with a long history in mathematics, computer science and engineering. Nevertheless, applying these approaches to neuroscience requires adaptation of their methods to the specific challenges of neural data. There have been a number of recent advances in this area, and we anticipate further advancement as part of the theoretical research supported by the SCGB. A still unresolved issue in applying maximum likelihood methods to neural data is that the basic statistics of spiking is Poisson whereas classic methods primarily apply Gaussian statistics. Classic dimensionality-reduction methods such as principal components analysis (PCA) have, in recent years, been extended and specialized for neuroscience applications. Thus our ability to apply statistical methods to neuroscience data is expanding rapidly, particularly in our ability to infer information and brain states in recordings from neural populations. Nevertheless, we are only beginning to explore newer ideas in the mathematics and statistics of high-dimensional spaces that are likely to be important in interpreting brain coding and dynamics. Indeed, it is vital to infuse new thinking and ideas into this field.
Similarly, physical scientists and engineers have long employed the tools of dynamical systems analysis to study the behavior of complex physical systems, but the constraints and special features of neural systems create challenges that need resolving. The primary difference between the dynamics of neural and other systems is that the information transmitted between neurons is typically in the ‘digital’ form of action potentials or spikes. This is a game changer in dynamical systems. There have been some significant recent advances in our ability to construct networks of spiking neurons that perform interesting tasks. Such models serve two purposes in the overall goals of the Simons collaboration. First, they allow us to apply the analytical tools discussed in the previous paragraph to realistic spike trains from models to see if they are informative about the underlying mechanisms, which are known in these models. Second, they provide a framework for testing hypotheses about how neural circuits operate.
In the past, research on neural networks has focused on static tasks involving the processing of sensory information such as object recognition and categorization, often using feed-forward connectivity. Real brain circuits are connected in a highly recurrent manner, and it is their dynamics, not their static properties, that link sensory representation to motor action. Recent advances in our ability to build and analyze recurrent networks allow us to construct models that duplicate the heterogeneity and variability of real neural circuits. Although these models are complex, we are beginning to have the tools to understand how they work and, in collaboration with experimentalists, to propose tests, not to determine if the model is ‘right’ but to answer the more interesting question of whether the mechanisms being used by the model to solve a task are those that the real neural circuits employ. We can also use our understanding of the model to reveal the biophysical features of the circuit, both cellular and synaptic, and the important aspects of its connectivity that give rise to its mode of operation. Again, these models suggest experimental tests, and our ability to manipulate real circuits, through optogenetic methods in particular, provides us with a path to extend results about general mechanisms to the level of specific circuit elements and features.
Modeling is done at a range of levels extending from conceptual to biophysical. For example, the application of drift-diffusion models to decision making ranges from a purely phenomenological description of behavior to network models of sensory integration. Similarly, recent studies of population activity in the prefrontal cortex range from using readout techniques to infer how information is represented to network simulations that model the dynamics of task switching. Future explorations of brain dynamics and cognition are likely to require a range of models, and this range will be supported within the SCGB.
In summary, developments in theory must keep pace with the remarkable ongoing technical advances in recording methods if we are to extract understanding from data. The collaboration is poised to contribute in a unique way to the enhancement as well as the application of data analysis and modeling methods that are vital to its overall goals.
3. Principles and Systems
The guiding principle of SCGB is to combine the latest innovative neurotechnology for recording and stimulating neural populations with the most powerful forms of analysis and modeling. This marriage of theory and experiment at the highest level will be used to first explore and ultimately answer the deep intellectual questions about how the dynamics of neural circuits produce cognition. The ambitions of this research program are substantial. We aim to provoke fundamental change in the way high-level brain states and internal dynamics are studied and understood to emerge from the activity of their elemental components. From a bottom-up perspective, we wish to know how neural computations within small circuits provide building blocks for synthetic operations at higher levels. From a top-down perspective, we seek insight into how high-level circuitry orchestrates the activity of low-level mechanisms in pursuit of larger behavioral goals. In short, we aim for an integrated understanding of nervous system function that takes mechanism seriously while avoiding a dull reductionism that mistakes a parts list for understanding. Powerful new analytical methods applied to multi-neuron recordings will ultimately permit studies of activity deep in the brain, remote from tight sensory or motor correlations, and will give meaningful information about internally generated, evanescent patterns of activity.
Our focus is on questions, not on particular experimental preparations or techniques. We aim to support research using whatever approach is most likely to advance our understanding. The ultimate aim is a mechanistic understanding of how cognitive processes are performed in human brains. Humans provide unparalleled flexibility in behavior, but experimentation on humans is extremely restricted and often not at the level of cellular-resolution population coding that is the main focus of this collaboration. Nonhuman primates have brains similar to those of humans and can be trained to do sophisticated tasks. Rats are obviously more distant from humans and more limited in the tasks they can perform, but they allow large numbers of subjects to be studied using a range of manipulations beyond what can be done in primates. Mice are more limited still, but the power of mouse genetics introduces another broad set of manipulations. Animals with small brains, such as the larval zebra fish, allow for whole-brain imaging of activity. Flies have both the advantages of a small nervous system and of extremely well developed methods for genetic manipulation. Worms add a known connectome to the list. Thus there exists a wide range of experimental systems, each with distinct advantages and disadvantages, for exploring the questions we raise. The SCGB will use these experimental systems and models of them to investigate the general principles of neural coding and dynamics and how they produce cognition and behavior.
The SCGB is directed by David W. Tank, Ph.D.
Henry L. Hillman Professor in Molecular Biology and Professor of Molecular Biology, Princeton University
Co-Director, Princeton Neuroscience Institute.
He chairs an executive committee that comprises the following individuals:
William Bloor Professor of Theoretical Neuroscience, Departments of Neuroscience, Physiology & Cellular Biophysics, and Biological Sciences, Columbia University
Co-Director, Center for Theoretical Neuroscience, and Member, The Kavli Institute for Brain Science, Columbia University
The Wilhelmina Robertson Professor and chair of the Department of Neuroscience, Baylor College of Medicine, with a joint appointment in the Departments of Electrical and Computer Engineering and Psychology, Rice University.
Associate professor of physiology and biophysics at the University of Washington, director of the university’s computational neuroscience program, and co-director of its Institute for Neuroengineering.
University Professor, Silver Professor, and Director, Center for Neural Science, New York University
Harman Family Provostial Professor, Director of Stanford Neurosciences Institute, and Professor of Neurobiology, Stanford University School of Medicine
Howard Hughes Medical Institute Investigator, Stanford University
Chief Scientist and Fellow of the Simons Foundation
A key component of the SCBG is the interaction among the members. Awardees are expected to share data and information and participate in large and small conferences and meetings organized by the Simons Foundation.
Future opportunities will include postdoc grant awards.