Abstracts

Day Two — Friday, August 24, 2012

Session 1: Keynote Lectures

Neural Coding: Building Brain Observatories at the Allen Institute
Christof Koch, PhD, Allen Institute for Brain Science

The Allen Institute for Brain Science is initiating a ten-year project to study the principles by which information is encoded, transformed and represented in the mammalian cerebral cortex and related structures. The Institute will build a series of brain observatories to identify, record and intervene in the neuronal networks underlying visually guided behaviors in the mouse, including visual perception, decision making and consciousness. This is a large-scale, in-house team effort to synthesize anatomical, physiological and theoretical knowledge into a description of the wiring scheme of the cortex, at both the structural and the functional levels. The fruits of this cerebroscope will be freely available to the public.

Mapping Gene Expression and Connections in the CNS: Tools and Data from the Allen Institute for Brain Science
Allan Jones, PhD, Allen Institute for Brain Science

The Allen Institute for Brain Science is a non-profit research organization dedicated to providing tools and data for the larger research community. Since 2003, the Allen Institute has created a suite of large-scale data efforts along with a web portal to view and analyze the data. These efforts include gene expression atlases of the developing and adult mouse brain and spinal cord, developing and adult human and non-human primate gene expression studies, and more recent efforts on connectivity atlases of the mouse brain. This presentation will cover an overview of the Allen Institute, its current projects and infrastructure, a few data highlights, and a look at future directions.

The Human Macro-connectome
David Van Essen, PhD, Washington University in St. Louis

Recent advances in noninvasive neuroimaging have set the stage for the systematic exploration of human brain circuits in health and disease. One such effort is the Human Connectome Project (HCP), which will characterize brain circuitry and its variability in healthy adults. A consortium of investigators at Washington University, University of Minnesota, University of Oxford, and 7 other institutions is engaged in a 5-year project to characterize the human connectome in 1,200 individuals (twins and their non-twin siblings). Information about structural and functional connectivity will be acquired using diffusion MRI and resting-state fMRI, respectively. Additional modalities will include task-evoked fMRI and MEG/EEG, plus extensive behavioral testing and genotyping. Advanced visualization and analysis methods will enable characterization of brain circuits in individuals and group averages at high spatial resolution and at the level of functionally distinct brain parcels (cortical areas and subcortical nuclei). Comparisons across subjects will reveal aspects of brain circuitry which are related to particular behavioral capacities and which are heritable or related to specific genetic variants.
 
Data from the HCP will be made freely available to the neuroscience community. A user-friendly informatics platform will enable investigators around the world to carry out many types of data mining on these freely accessible, information-rich datasets. Altogether, the HCP will provide invaluable information about the healthy human brain and its variability.

Blue Brain: Insights From the Synthesis of a Cortical Column
Sean Hill, PhD, Ecole Polytechnique Federale de Lausanne

The Blue Brain Project aims to provide a generic facility for large-scale neuroscience data integration, modeling and simulation. A prototype facility has been completed, which is capable today of building neural microcircuits or modules of the rat brain with cellular level resolution. This prototype was founded on a novel data-driven and data-constrained process for creating, validating and researching the neocortical column. Recent models recreate key experimental findings of structural and functional properties of neocortical circuitry in vitro — including connectivity, synaptic responses and network dynamics. We present insights gained from this process including principles underlying invariance and robustness in cortical microcircuitry.

Reading and Writing All Basepairs in a Genome and All Impulses in a Brain
George Church, PhD, Harvard Medical School

We have brought down the cost of reading and writing DNA (in genomes and epigenomes) by about a million-fold in the past 8 years. Higher quality and comprehensiveness have greatly improved genomic models, software, and applications. Synthetic biology plus the decreasing size of components in wireless circuits will likely enable analogous exponential progress in brain-computer interfaces. The ability to inexpensively record, hyperpolarize, and depolarize precise combinations of neurons on demand is likely to be quite helpful for rapidly generating and testing models (see also PMID: 22726828).

Session 2: Advances in Tools, Technology, and Methodology: Innovative Toolbuilding, Neuroimaging, and Neuroinformatics

New Tools for Analyzing and Engineering Brain Circuits
Ed Boyden, PhD, Massachusetts Institute of Technology

Understanding how neural circuits implement brain functions and how these computations go awry in brain disorders, is a top priority for neuroscience. Achieving this understanding will require new technologies. Over the last several years we have developed a rapidly-expanding suite of genetically-encoded reagents that, when expressed in specific neuron types in the nervous system, enable their electrical activities to be powerfully and precisely activated and silenced in response to pulses of light. I will briefly give an overview of the field, and then I will discuss a number of new tools for neural activation and silencing that we are developing, including new molecules with augmented amplitudes, improved safety profiles, novel color and light-sensitivity capabilities, and unique new capabilities. Second, we have begun to develop microfabricated and robotic hardware to enable complex and distributed neural circuits to be precisely controlled, and for the network-wide impact of a neural control event to be measured using distributed electrodes, fMRI, and automated intracellular neural recording. We explore how these tools can be used to enable systematic analysis of neural circuit functions in the fields of emotion, sensation, and movement, and in neurological and psychiatric disorders.

Sequencing the Connectome
Anthony Zador, MD, PhD, Cold Spring Harbor Laboratory

The brain is an extremely complex network, consisting of billions of neurons connected by trillions of synapses. The details of these connections—which neurons form synaptic connections with which other neurons—are crucial in determining brain function. Malformation of these connections during prenatal and early postnatal development can lead to mental retardation, autism or schizophrenia; loss of specific connections later in life is associated with neurodegenerative diseases such as Alzheimer's. We are developing an entirely novel approach based on high-throughput DNA sequencing technology. Sequencing technology has not previously been applied in the context to neural connectivity. The appeal of using sequencing is that it is fast and cheap. Moreover, like microprocessor technology, sequencing technology is improving exponentially. An efficient method for determining the brain's wiring diagram would provide a foundation for understanding how neural circuits compute and could transform neuroscience research.

Imaging Neuronal Activity in the Freely Moving Animal: From the Eye to the Cortex
Jason N. D. Kerr, PhD, Network Imaging Group, Max Planck Institute for Biological Cybernetics, Germany

Motivation underlies the performance of self-determined behavior and is fundamental to decision making, especially with regard to seeking food, mates, and avoiding peril. As many decision making based behaviors in rodents involve a combination of head movements, eye movements, vestibular driven neuronal activity and multimodal active sensing of the environment to guide the behavior, studying the freely moving animal is paramount. To achieve this, what is also necessary is the precise tracking of the animal's movement and interaction with the environment. Here I will outline work from our group that focuses on how freely moving rodents use their vision during decision making tasks and resulting cortical activity. I will introduce methods that allow accurate recording of neuronal activity from populations of cortical neurons, using multi-photon imaging techniques, while simultaneously tracking behavior, using eye and head tracking techniques, during decision making in the freely moving rodent. The second half of the presentation will focus on recent results from our lab showing how rodents have a distinct eye movement strategy that is of major evolutionary benefit.

New Approaches for Correlated LM and 3D EM Applied to MULTISCALE CHALLENGES: Bridging Gaps in Knowledge and Understanding
Mark H. Ellisman, PhD, The National Center for Microscopy and Imaging Research (NCMIR), University of California, San Diego

A grand goal in cell biology is to understand how the interplay of structural, chemical and electrical signals in and between cells gives rise to tissue properties, especially for complex tissues like nervous systems. New technologies are hastening progress as biologists make use of an increasingly powerful arsenal of tools and technologies for obtaining data, from the level of molecules to whole organs, and at the same time engage in the arduous and challenging process of adapting and assembling data at all scales of resolution and across disciplines into computerized databases. This talk will highlight projects in which development and application of new contrasting methods and imaging tools have allowed us to observe otherwise hidden relationships between cellular, subcellular and molecular constituents of cells, including those of nervous systems.
 
New chemistries for carrying out correlated light and electron microscopy will be described, as well as recent advances in large-scale high-resolution 3D reconstruction with LM, TEM and SEM based methods. Examples of next generation cell-centric image libraries and web-based multiscale information exploration environments for sharing and exploring these data will also be described.

Developing an International Neuroinformatics Infrastructure
Sean Hill, PhD, Karolinska Institute

The International Neuroinformatics Coordinating Facility (INCF) was launched in 2005, following the proposal by the Global Science Forum of the Organization for Economic Cooperation and Development (OECD) to create an organization to coordinate an open international infrastructure to integrate heterogeneous neuroscience data and knowledge bases and enable new insights from analysis, modeling and simulation. Here we present the INCF multi-phase strategy to deploy such an infrastructure with specific capabilities and milestones. The first phase is to establish a globally federated data space with searchable metadata. The second phase will deploy an object-based data integration layer employing web services to ensure the unique identification of all data through ontologies and spatial coordinates, while using data models to access diverse data formats through standard interfaces. The third phase would enable standard workflow management for analysis, visualization, modeling and simulation can then be built on top of the data integration layer. The development of portal interfaces will be critical to provide interactive user access to data, analyses and simulation results. The aim of this infrastructure is to facilitate international sharing, publication and integration of neuroscience data across multiple levels and scales from genes to behavior.

Session 3: Advances in Tools, Technology and Methodology: Computational Models

Prediction in the Retina
Stephanie E. Palmer, PhD, University of Chicago

In the natural world, temporal correlations between events exist on many timescales, allowing organisms to anticipate the future state of their environments. A neural system that uses predictions to guide behavior must encode the future values of sensory inputs. This suggests a new approach to neural encoding. While most studies have, historically, sought to characterize what stimuli in the past gave rise to a response (the classical receptive field picture of encoding), we ask instead what stimuli those responses predict. We have found such 'predictive information' in the population responses of retinal ganglion cells (RGCs) in the larval salamander. To quantify predictive information, we ask how much RGC responses at some time 'now' (R_p) tell us about the future state of the stimulus (S_f). This information, I(R_p;S_f), is bounded by correlations in the stimulus itself, I(S_p;S_f). For particular classes of stimuli, this bound can be calculated analytically. We have shown that certain patterns of population firing in the retina approach this bound, suggesting that the retina may be optimized for prediction. Coding for prediction may be a useful strategy for neural systems to adopt, making transfer of sensory information more efficient by compressing signals along dimensions relevant for behavior.

Coauthors: Olivier Marre, PhD², Michael J. Berry, II, PhD³, William Bialek, PhD^3
2Paris VI University, Paris, France
³Princeton University, Princeton, NJ

Bayesian Inference with Efficient Neural Population Codes
Alan A. Stocker, PhD, University of Pennsylvania

The accuracy with which the perceptual brain can infer the value of stimulus variables in the worlddepends on both, the amount of stimulus information that is represented in a population of sensory neurons (encoding) and the mechanism by which this information is subsequently retrieved from the population's response pattern (decoding).
 
Previous studies have mainly focused either on the encoding or the decoding aspect of the problem, providing evidence for two general optimality principles: The efficient coding hypothesis (Barlow 1961) states that neural representations are optimally adapted to encode a given stimulus ensemble, while the Bayesian hypothesis proposes that the brain is able to optimally decode a stimulus variable by combining sensory evidence with prior information (e.g. Knill/Richards 1999).
 
Here I present recent work of my laboratory in which we developed a new theoretical framework that functionally links optimal (efficient) encoding with optimal (Bayesian) decoding. More specifically, I demonstrate that efficient population codes allow the accurate emulation of Bayesian inference with a relative simple, neural decoding mechanism based on a generalized form of the population vector read-out. Stimulus priors (bottom-up) are intrinsically represented by the tuning curves distribution in the neural population, while top-down attentional priors can be incorporated by gain changes in neural firing. The framework makes specific predictions about perceptual behavior based on stimulus specific parameters such as stimulus prior, strength and time-constants, as well as physiological parameters such as spontaneous firing rates. The framework is a concrete example for the duality between neural representation and computation.

The Orchestral Brain: Coding with Correlated and Heterogeneous Neurons
Rava Azeredo da Silveira, PhD, Ecole Normal Supérieure, Paris

While single-cell activity may be well correlated with simple aspects of sensory stumuli, rich stimuli or subtly differing stimuli require concomitant coding by several neurons in a population. It is then natural to ask whether the nature of the coding is 'orchestral' in that it relies upon correlation and physiological diversity among cells. Positive correlations in the activity of neurons are widely observed in the brain and previous studies stipulate that these are at best marginally favorable, if not detrimental, to the fidelity of population codes, compared to independent codes. Here, we put forth a scenario in which positive correlations can enhance coding performance by astronomical factors. Specifically, the probability of discrimination error can be suppressed by many orders of magnitude. Likewise, the number of stimuli encoded—the capacity—can be enhanced by similarly large factors. These effects do not necessitate unrealistic correlation values and can occur for populations with as little as a few tens of neurons. The scenario relies upon 'lock-in' patterns of activity with which correlation relegates the noise in irrelevant modes. We further demonstrate that, quite generically, coding fidelity is enhanced by physiological heterogeneity. Finally, we formulate heuristic arguments as to the plausibility of 'lock-in' patterns and possible experimental tests of the theoretical proposal.

View from the Top: What Probabilistic Models of Perception Can Teach Us about Neural Computation
Wei Ji Ma, PhD, Baylor College of Medicine

Sensory information is often noisy and ambiguous and perception is uncertain as a result. Under such circumstances, organisms can maximize their performance by using a decision strategy known as probabilistic or Bayesian inference. In simple perceptual tasks such as cue combination, Bayesian models describe human behavior extremely well. Here, we show that the formalism extends to more cognitive tasks, where inference is typically categorical and hierarchical, and resource limitations might play a role. As examples, we will discuss visual search, change detection, and categorization under ambiguity. Probabilistic models of perceptual and cognitive behavior provide strong constraints on theories of the underlying neural computations and yield testable predictions for physiological experiments. We will illustrate this using cue combination, a realm in which these physiological predictions have partially been confirmed.

Computing Intelligence: Mind, Brain and Machine
Tomaso Poggio, PhD, Massachusetts Institute of Technology

I conjecture that the sample complexity of object recognition is mostly due to geometric image transformations and that a main goal of the ventral stream is to learn-and-discount image transformations. The theory predicts that the size of the receptive fields determines which transformations are learned during development; that the transformation represented in each area determines the tuning of the neurons in the area; and that class specific transformations are learned and represented at the top of the ventral stream hierarchy. If the theory were true, the ventral system would be a mirror of the symmetry properties of motions in the physical world.