Engineering

In this chapter, our objective is to provide the reader a brief introduction to learning and memory as viewed from the perspective of neural computation. The  computational agenda in cognitive neuroscience focuses on two important tasks, each relevant to the study of learning and memory: to identify and understand how neural computations are carried out in the brain and to use computational simulations to better understand how these neural computations give rise to cognitive function. Our agenda is (i) to present a set of core principles of neural computation that have been derived from empirical work in neuroscience and (ii) to demonstrate how these principles can serve as building blocks for computational models that have been used to explain many aspects of learning and memory.

We review evidence that in the course of reading, the visual system computes abstract letter identities (ALIs): a representation of letters that encodes their identity but that abstracts away from their visual appearance. How could the visual system learn such a seemingly nonvisual representation? We propose that different forms of the same letter tend to appear in similar distributions of contexts (in the same words written in different ways) and that this environmental correlation interacts with correlation-based learning mechanisms in the brain to lead to the formation of ALIs. We review a
neural network model that demonstrates the feasibility of this common contexts hypothesis and present two experiments confirming some novel predictions: (a) repeatedly presenting arbitrary visual stimuli in common contexts leads those stimuli to be confusable with each other, and (b) different forms of the same letter are more confusable with each other in word-like contexts than in nonword-like contexts. We then extend the model to use real pictures of letters as input and simulate some of the novel empirical findings from the experiments.

A number of behavioral and neuroscientific studies suggest that face processing is qualitatively different from the processing of other visual stimuli. Why? Is face processing in some sense innate? What role does experience play in the development of face processing? We review recent evidence related to these questions. We begin by identifying some of the ways in which face processing is special. We then consider findings that demonstrate a crucial role for experience-independent genetic mechanisms in the development of face processing and its neural substrates. Finally, we review studies demonstrating the crucial role played by experience-dependent mechanisms. These findings support the hypothesis that there is a genetic predisposition for a special face processing mechanism, but that experience plays a crucial role in tuning this mechanism during development.

Neoclassical models of strategic behavior have yielded many insights into competitive behavior, despite the fact that they often rely on a number of assumptions—including instantaneous market clearing and perfect foresight—that have been called into question by a broad range of research. Researchers generally argue that these assumptions are "good enough" to predict an industry’s probable equilibria, and that disequilibrium adjustments and bounded rationality have limited competitive implications. Here we focus on the case of strategy in the presence of increasing returns to highlight how relaxing these two assumptions can lead to outcomes quite different from those predicted by standard neoclassical models. Prior research suggests that in the presence of increasing returns, tight appropriability, and accommodating rivals, in some circumstances early entrants can achieve sustained competitive advantage by pursuing "get big fast" (GBF) strategies: Rapidly expanding capacity and cutting prices to gain market share advantage and exploit positive feedbacks faster than their rivals. Using a simulation of the duopoly case we show that when the industry moves slowly compared to capacity adjustment delays, boundedly rational firms find their way to the equilibria predicted by conventional models. However, when market dynamics are rapid relative to capacity adjustment, forecasting errors lead to excess capacity—overwhelming the advantage conferred by increasing returns. Our results highlight the risks of ignoring the role of disequilibrium dynamics and bounded rationality in shaping competitive outcomes, and demonstrate how both can be incorporated into strategic analysis to form a dynamic, behavioral game theory amenable to rigorous analysis

A number of findings suggest that genetics play a more important role in shaping the processing of faces compared with other visual stimuli. For example, brain damage immediately after birthcan produce lasting, selective deficits in face recognition, but we know of no similar reports of early brain damage leading to selective deficits in the recognition of other categories like animals or furniture. Patterns of cortical activation evoked by face stimuli are also more similar in
monozygotic twins than in dizygotic twins, but this is not true for patterns evoked by other visual stimuli such as words and chairs. By what mechanism could genetics come to influence the cortical development of face processing more than the processing of other visual stimuli? We hypothesize that an innate attentional preference to look at faces in early development interacts with initial functional connectivity and principles of cortical self-organization to lead to observed
neural and behavioral effects. We present a neural network model that makes explicit how this hypothesis might work. In particular, we demonstrate that manipulating the similarity of initial patterns of connectivity in simulated cortex has a greater effect on the similarity of activity patterns evoked by faces than by other visual stimuli. The model therefore provides a computationally explicit account of the observed effects of zygosity on neural activation patterns in human twins and, more generally, explains how genetics could influence face processing preferentially.

The Iowa Gambling Task (IGT) is a well–studied experimental paradigm known to simulate both intact and impaired real-world decision making in choice tasks that involve uncertain payoffs. Prior work has used computational reinforcement learning models to successfully reproduce a range of task phenomena. In this prior work a set of models were fit to individual decision making data, the best-fit models were selected based on group-averaged metrics, and then theoretical conclusions were drawn based on group-averaged parameters. In the present work we investigate the performance of this class of reinforcement learning models in fitting individual data. This class of learning models has provided a useful starting point for characterizing decision making performance in the aggregate. However, we demonstrate that no one best-fit model aptly captures individual differences and our results caution against using aggregate parameters from best fit models to characterize decision making across populations as has been done in prior work.

Representations in sensory cortices are organized topographically: auditory cortex is organized tonotopically, somatosensory cortex is organized somatotopically, and visual cortex is organized retinotopically. Substantial progress has been made in understanding how topography develops at a neurocomputational level, particularly in the early and middle stages of processing in the visual system. We extend this work to investigate how higher-level semantic representations could develop based on topographic input from sensory maps in the ventral visual pathway. The receptive fields of cells in these maps correspond to the loci of activity within a cortical topography rather than explicitly coded sensory features. Using this model, we show that meaningful semantic representations at increasing levels of abstraction naturally emerge as a result of exposure to a set of visual stimuli. For example, when presented with a set of simple visual features (color, texture, size, and shape) the model develops semantic representations that distinguish basic level categories (e.g., dogs, tables, cars), superordinate categories (e.g., animals, furniture, vehicles), and living versus nonliving things. This work therefore offers a computationally explicit hypothesis about how semantic representations could emerge in the brain. Our results suggest the possibility that high-order concept representations may be encoded topographically much the same way as low-order sensory representations, and that these representations may be learned based on the same principles of neural computation known to be operating in sensory cortex.

Decision making, the process of choosing among a set of options, is a fundamental aspect of everyday mental life.  Decisions are often made under conditions of uncertainty, when the payoffs are probabilistic and unknown.  Two important challenges in the study of decision making are to understand how decision making processes are instantiated computationally and to reveal and characterize differences in decision making processes across individuals.  In this dissertation I used computational and behavioral methods to study decision making under uncertainty in the context of the Iowa Gambling Task (IGT).  The main contributions are:  (i) A biologically-grounded computational model that provides a better account of IGT behavior than the most widely accepted model; (ii) The identification of three fundamentally different decision making styles in the IGT; (iii) An improved conceptualization of decision making that offers a more comprehensive approach to analyzing performance and that has important implications for the study of normal and clinically impaired decision making; (iv) An empirical challenge to the widely held belief that IGT performance is associated with impulsive and risky decision making;  (v) A demonstrated association between decision making in the IGT and cognitive abilities as measured by the Wisconsin Card Sorting Task; and (vi) the introduction and successful demonstration of a robust data clustering methodology that offers significant advantages over methods that are currently used in the field of Psychology.

This paper experimentally investigates the performance of multiple-input multiple-output (MIMO) systems in indoor ultra wideband channels. The improvement in robustness and information rate due to spatial and polar antenna arrays is evaluated. The subchannel correlation, power gains of supported eigenmodes, and branch power ratios are analyzed. The polar arrays are found to experience lower correlation than spatial arrays. SNR gains of up to 3–5 dB are reported with 1x2 and 1x3 spatial arrays, respectively; the latter is shown to double the coverage range. The mutual information capacity is found to scale almost linearly with the MIMO array size, with very low variance. It is confirmed that the device compactness achieved by the polar array comes with only a small penalty in the achievable capacity and SNR gain compared to the spatial array. The multiple-antenna UWB techniques explored in this paper offer the potential for high data rate and robust communications.

Correlated fading adversely affects the spatial multiplexing and diversity performance of multiple-antenna systems. This paper presents an empirical investigation of spatial correlation in ultrawideband (UWB) indoor channels in the FCC- allocated 3.1-10.6 GHz band. The impact of a number of system parameters on spatial correlation at the receiver is evaluated. It is found that the coherence distance falls with channel bandwidth in endfire arrays but not in broadside arrays. The complex correlation decays less rapidly with distance in broadside arrays than in endfire arrays, especially under line-of-sight, but the envelope correlation is circularly symmetric in space. Strong dependence of spatial correlation and coherence distance on the channel center frequency is observed. Furthermore, horizontally polarized links are found to exhibit less correlation than vertically polarized links. A discussion of the implications of these results for multiple-antenna UWB systems is provided.

The variation of channel fading and dispersion statistics between the ultrawideband (UWB) channel subbands is characterized experimentally. Within the FCC-allocated UWB frequency range (3.1-10.6 GHz), the multipath characteristics in various subbands are found to be highly dissimilar. It is demonstrated that the higher subbands suffer greater attenuation, and the rms delay spread varies significantly with the center frequency. The impact of this variation on multiband-OFDM UWB system performance is evaluated, and a large performance variation across the band is reported.

The polarization characteristics of the indoor ultrawideband channel are investigated using dual-polar 3.1-10.6 GHz channel measurements. Under the generalized elliptical polarization framework, the channel cross-polar discrimination (XPD) statistics are evaluated. Lognormal and triangular distributions are shown to well model the XPD and ellipticity angle. Considerable spectral and small-scale spatial variability is observed in the XPD. Asymptotic convergence of XPD to its mean with increasing bandwidth is established. An analysis of the variation of tap XPD with excess delay indicates that the XPD is highest for the initial paths. It is shown that increased scattering leads to greater depolarization. The dependence of XPD on the channel's rms delay spread is modelled using a linear regression with -0.5 dB/ns slope. This characterization will aid the analysis of polarization-dependent loss and the design of polarized multiple-antenna systems.

Rake reception can improve system performance significantly in wideband multipath channels. Its practical implementation, however, becomes prohibitively expensive in channels with dense multipath, such as the ultrawideband (UWB) channel. This paper investigates the effect of various system and environment parameters on rake performance, with emphasis on the amount of multipath and channel bandwidth. The treatment includes hybrid selection/maximal-ratio combining (H-S/MRC) and unordered, partial combining rake architectures, and is based on indoor channel measurements in the FCC-allocated UWB frequency range (3.1-10.6 GHz). The diversity gain is shown to follow the law of diminishing returns with the rake complexity. It is demonstrated that the rake can extract most of the incident signal power by combining only a subset of the resolved multipath components. The required number of rake fingers increases linearly with the number of resolved paths but sublinearly with the channel bandwidth. The characterization of the interplay of bandwidth, amount of scattering and rake complexity will facilitate efficient implementation of UWB systems.

The radiation characteristics of ultrawideband (UWB) antennas vary with frequency, introducing directionally asymmetric bandwidth reduction and waveform dispersion. In this paper, we develop a simple technique to alleviate the distortion due to nonisotropically dispersive antennas, and use indoor channel measurements to verify its performance. The approach is based on multipath direction estimation and therefore involves antenna arrays. We show that antenna distortion can enhance sensor localization ambiguity and introduce errors in its estimate. Antenna compensation mitigates this effect, significantly improving the location estimation accuracy. We further demonstrate that antenna compensation helps reduce the small-scale fading artifacts that arise due to the antennas, thus reducing the channel spatial variability and delay spread. Our technique can also aid empirical channel characterization by providing antenna-independent propagation data.

The Kalman filter is commonly used in neural interface systems to decode neural activity and estimate the desired movement kinematics. We analyze a low-complexity Kalman filter implementation in which the filter gain is approximated by its steady-state form, computed offline before real-time decoding commences. We evaluate its performance using human motor cortical spike train data obtained from an intracortical recording array as part of an ongoing pilot clinical trial. We demonstrate that the standard Kalman filter gain converges to within 95% of the steady-state filter gain in 1.5 ± 0.5 s (mean ±s.d.). The difference in the intended movement velocity decoded by the two filters vanishes within 5 s, with a correlation coefficient of 0.99 between the two decoded velocities over the session length. We also find that the steady-state Kalman filter reduces the computational load (algorithm execution time) for decoding the firing rates of 25±3 single units by a factor of 7.0±0.9. We expect that the gain in computational efficiency will be much higher in systems with larger neural ensembles. The steady-state filter can thus provide substantial runtime efficiency at little cost in terms of estimation accuracy. This far more efficient neural decoding approach will facilitate the practical implementation of future large-dimensional, multisignal neural interface systems.

Two-photon calcium imaging is now an important tool for in vivo imaging of biological systems. By enabling neuronal population imaging with subcellular resolution, this modality offers an approach for gaining a fundamental understanding of brain anatomy and physiology. Proper analysis of calcium imaging data requires denoising, that is separating the signal from complex physiological noise. To analyze two-photon brain imaging data, we present a signal plus colored noise model in which the signal is represented as harmonic regression and the correlated noise is represented as an order p autoregressive process. We provide an efficient cyclic descent algorithm to compute approximate maximum likelihood parameter estimates by combing a weighted least-squares procedure with the Burg algorithm. We use Akaike information criterion to guide selection of the harmonic regression and the autoregressive model orders. Our flexible yet parsimonious modeling approach reliably separates stimulus-evoked fluorescence response from background activity and noise, assesses goodness of fit, and estimates confidence intervals and signal-to-noise ratio. This refined separation leads to appreciably enhanced image contrast for individual cells including clear delineation of subcellular details and network activity. The application of our approach to in vivo imaging data recorded in the ferret primary visual cortex demonstrates that our method yields substantially denoised signal estimates. We also provide a general Volterra series framework for deriving this and other signal plus correlated noise models for imaging. This approach to analyzing two-photon calcium imaging data may be readily adapted to other computational biology problems which apply correlated noise models.

Over the last decade there has been a tremendous advance in the analytical tools available to neuroscientists to understand and model neural function. In particular, the point process – generalized linear model (PP-GLM) framework has been applied successfully to problems ranging from neuro-endocrine physiology to neural decoding. However, the lack of freely distributed software implementations of published PP-GLM algorithms together with problem-specific modifications required for their use, limit wide application of these techniques. In an effort to make existing PP-GLM methods more accessible to the neuroscience community, we have developed nSTAT – an open source neural spike train analysis toolbox for Matlab®. By adopting an object-oriented programming (OOP) approach, nSTAT allows users to easily manipulate data by performing operations on objects that have an intuitive connection to the experiment (spike trains, covariates, etc.), rather than by dealing with data in vector/matrix form. The algorithms implemented within nSTAT address a number of common problems including computation of peri-stimulus time histograms, quantification of the temporal response properties of neurons, and characterization of neural plasticity within and across trials. nSTAT provides a starting point for exploratory data analysis, allows for simple and systematic building and testing of point process models, and for decoding of stimulus variables based on point process models of neural function. By providing an open-source toolbox, we hope to establish a platform that can be easily used, modified, and extended by the scientific community to address limitations of current techniques and to extend available techniques to more complex problems.► We have developed the neural spike train analysis toolbox (nSTAT) for Matlab®. ► nSTAT makes existing point process/GLM methods for spike train analysis more accessible to the neuroscience community. ► nSTAT adopts object-oriented programming to allow manipulation of data objects rather than raw numerical representations. ► nSTAT allows systematic building/testing of neural encoding models and allows these models to be used for neural decoding.

Abstract In most wireless communications research, the channel models considered experience less severe fading than the classic Rayleigh fading case. In this work, however, we investigate MIMO channels where the fading is more severe. In these environments, we show that the coefficient of variation of the channel amplitudes is a good predictor of the link mutual information, for a variety of models. We propose a novel channel model for severely fading channels based on the complex multivariate t distribution.