Exploiting semantic information in a spiking neural SLAM system.

To navigate in new environments, an animal must be able to keep track of its position while simultaneously creating and updating an internal map of features in the environment, a problem formulated as simultaneous localization and mapping (SLAM) in the field of robotics. This requires integrating information from different domains, including self-motion cues, sensory, and semantic information. Several specialized neuron classes have been identified in the mammalian brain as being involved in solving SLAM. While biology has inspired a whole class of SLAM algorithms, the use of semantic information has not been explored in such work. We present a novel, biologically plausible SLAM model called SSP-SLAM-a spiking neural network designed using tools for large scale cognitive modeling. Our model uses a vector representation of continuous spatial maps, which can be encoded via spiking neural activity and bound with other features (continuous and discrete) to create compressed structures containing semantic information from multiple domains (e.g., spatial, temporal, visual, conceptual). We demonstrate that the dynamics of these representations can be implemented with a hybrid oscillatory-interference and continuous attractor network of head direction cells. The estimated self-position from this network is used to learn an associative memory between semantically encoded landmarks and their positions, i.e., an environment map, which is used for loop closure. Our experiments demonstrate that environment maps can be learned accurately and their use greatly improves self-position estimation. Furthermore, grid cells, place cells, and object vector cells are observed by this model. We also run our path integrator network on the NengoLoihi neuromorphic emulator to demonstrate feasibility for a full neuromorphic implementation for energy efficient SLAM.

Biological batch normalisation: How intrinsic plasticity improves learning in deep neural networks.

In this work, we present a local intrinsic rule that we developed, dubbed IP, inspired by the Infomax rule. Like Infomax, this rule works by controlling the gain and bias of a neuron to regulate its rate of fire. We discuss the biological plausibility of the IP rule and compare it to batch normalisation. We demonstrate that the IP rule improves learning in deep networks, and provides networks with considerable robustness to increases in synaptic learning rates. We also sample the error gradients during learning and show that the IP rule substantially increases the size of the gradients over the course of learning. This suggests that the IP rule solves the vanishing gradient problem. Supplementary analysis is provided to derive the equilibrium solutions that the neuronal gain and bias converge to using our IP rule. An analysis demonstrates that the IP rule results in neuronal information potential similar to that of Infomax, when tested on a fixed input distribution. We also show that batch normalisation also improves information potential, suggesting that this may be a cause for the efficacy of batch normalisation-an open problem at the time of this writing.

A Predictive-Coding Network That Is Both Discriminative and Generative.

Predictive coding (PC) networks are a biologically interesting class of neural networks. Their layered hierarchy mimics the reciprocal connectivity pattern observed in the mammalian cortex, and they can be trained using local learning rules that approximate backpropagation (Bogacz, 2017). However, despite having feedback connections that enable information to flow down the network hierarchy, discriminative PC networks are not typically generative. Clamping the output class and running the network to equilibrium yields an input sample that usually does not resemble the training input. This letter studies this phenomenon and proposes a simple solution that promotes the generation of input samples that resemble the training inputs. Simple decay, a technique already in wide use in neural networks, pushes the PC network toward a unique minimum two-norm solution, and that unique solution provably (for linear networks) matches the training inputs. The method also vastly improves the samples generated for nonlinear networks, as we demonstrate on MNIST.

PixelBNN: Augmenting the PixelCNN with Batch Normalization and the Presentation of a Fast Architecture for Retinal Vessel Segmentation.

Analysis of retinal fundus images is essential for eye-care physicians in the diagnosis, care and treatment of patients. Accurate fundus and/or retinal vessel maps give rise to longitudinal studies able to utilize multimedia image registration and disease/condition status measurements, as well as applications in surgery preparation and biometrics. The segmentation of retinal morphology has numerous applications in assessing ophthalmologic and cardiovascular disease pathologies. Computer-aided segmentation of the vasculature has proven to be a challenge, mainly due to inconsistencies such as noise and variations in hue and brightness that can greatly reduce the quality of fundus images. The goal of this work is to collate different key performance indicators (KPIs) and state-of-the-art methods applied to this task, frame computational efficiency-performance trade-offs under varying degrees of information loss using common datasets, and introduce PixelBNN, a highly efficient deep method for automating the segmentation of fundus morphologies. The model was trained, tested and cross tested on the DRIVE, STARE and CHASE_DB1 retinal vessel segmentation datasets. Performance was evaluated using G-mean, Mathews Correlation Coefficient and F1-score, with the main success measure being computation speed. The network was 8.5× faster than the current state-of-the-art at test time and performed comparatively well, considering a 5× to 19× reduction in information from resizing images during preprocessing.

Symmetric Predictive Estimator for Biologically Plausible Neural Learning.

In a real brain, the act of perception is a bidirectional process, depending on both feedforward sensory pathways and feedback pathways that carry expectations. We are interested in how such a neural network might emerge from a biologically plausible learning rule. Other neural network learning methods either only apply to feedforward networks, or employ assumptions (such as weight copying) that render them unlikely in a real brain. Predictive estimators (PEs) offer a better solution to this bidirectional learning scenario. However, PEs also depend on weight copying. In this paper, we propose the symmetric PE (SPE), an architecture that can learn both feedforward and feedback connection weights individually using only locally available information. We demonstrate that the SPE can learn complicated mappings without the use of weight copying. The SPE networks also show promise in deeper architectures.

Improving Neural-Network Classifiers Using Nearest Neighbor Partitioning.

This paper presents a nearest neighbor partitioning method designed to improve the performance of a neural-network classifier. For neural-network classifiers, usually the number, positions, and labels of centroids are fixed in partition space before training. However, that approach limits the search for potential neural networks during optimization; the quality of a neural network classifier is based on how clear the decision boundaries are between classes. Although attempts have been made to generate floating centroids automatically, these methods still tend to generate sphere-like partitions and cannot produce flexible decision boundaries. We propose the use of nearest neighbor classification in conjunction with a neural-network classifier. Instead of being bound by sphere-like boundaries (such as the case with centroid-based methods), the flexibility of nearest neighbors increases the chance of finding potential neural networks that have arbitrarily shaped boundaries in partition space. Experimental results demonstrate that the proposed method exhibits superior performance on accuracy and average f-measure.

Oscillator-interference models of path integration do not require theta oscillations.

Navigation and path integration in rodents seems to involve place cells, grid cells, and theta oscillations (4-12 Hz) in the local field potential. Two main theories have been proposed to explain the neurological underpinnings of how these phenomena relate to navigation and to each other. Attractor network (AN) models revolve around the idea that local excitation and long-range inhibition connectivity can spontaneously generate grid-cell-like activity patterns. Oscillator interference (OI) models propose that spatial patterns of activity are caused by the interference patterns between neural oscillators. In rats, these oscillators have a frequency close to the theta frequency. Recent studies have shown that bats do not exhibit a theta cycle when they crawl, and yet they still have grid cells. This has been interpreted as a criticism of OI models. However, OI models do not require theta oscillations. We explain why the absence of theta oscillations does not contradict OI models and discuss how the two families of models might be distinguished experimentally.

Does the entorhinal cortex use the Fourier transform?

Some neurons in the entorhinal cortex (EC) fire bursts when the animal occupies locations organized in a hexagonal grid pattern in their spatial environment. Place cells have also been observed, firing bursts only when the animal occupies a particular region of the environment. Both of these types of cells exhibit theta-cycle modulation, firing bursts in the 4-12 Hz range. Grid cells fire bursts of action potentials that precess with respect to the theta cycle, a phenomenon dubbed "theta precession." Various models have been proposed to explain these phenomena, and how they relate to navigation. Among the most promising are the oscillator interference models. The bank-of-oscillators model proposed by Welday et al. (2011) exhibits all these features. However, their simulations are based on theoretical oscillators, and not implemented entirely with spiking neurons. We extend their work in a number of ways. First, we place the oscillators in a frequency domain and reformulate the model in terms of Fourier theory. Second, this perspective suggests a division of labor for implementing spatial maps: position vs. map layout. The animal's position is encoded in the phases of the oscillators, while the spatial map shape is encoded implicitly in the weights of the connections between the oscillators and the read-out nodes. Third, it reveals that the oscillator phases all need to conform to a linear relationship across the frequency domain. Fourth, we implement a partial model of the EC using spiking leaky integrate-and-fire (LIF) neurons. Fifth, we devise new coupling mechanisms, enlightened by the global phase constraint, and show they are capable of keeping spiking neural oscillators in consistent formation. Our model demonstrates place cells, grid cells, and phase precession. The Fourier model also gives direction for future investigations, such as integrating sensory feedback to combat drift, or explaining why grid cells exist at all.

Population coding in sparsely connected networks of noisy neurons.

This study examines the relationship between population coding and spatial connection statistics in networks of noisy neurons. Encoding of sensory information in the neocortex is thought to require coordinated neural populations, because individual cortical neurons respond to a wide range of stimuli, and exhibit highly variable spiking in response to repeated stimuli. Population coding is rooted in network structure, because cortical neurons receive information only from other neurons, and because the information they encode must be decoded by other neurons, if it is to affect behavior. However, population coding theory has often ignored network structure, or assumed discrete, fully connected populations (in contrast with the sparsely connected, continuous sheet of the cortex). In this study, we modeled a sheet of cortical neurons with sparse, primarily local connections, and found that a network with this structure could encode multiple internal state variables with high signal-to-noise ratio. However, we were unable to create high-fidelity networks by instantiating connections at random according to spatial connection probabilities. In our models, high-fidelity networks required additional structure, with higher cluster factors and correlations between the inputs to nearby neurons.

Registering a non-rigid multi-sensor ensemble of images.

The majority of image registration methods deal with registering only two images at a time. Recently, a clustering method that concurrently registers more than two multi-sensor images was proposed, dubbed ensemble clustering. In this paper, we apply the ensemble clustering method to a deformable registration scenario for the first time. Non-rigid deformation is implemented by a free-form deformation model based on B-splines with a regularization term. However, the increased degrees of freedom in the transformations caused the Newton-type optimization process to become ill-conditioned. This made the registration process unstable. We solved this problem by using the matrix approximation afforded by the singular value decomposition (SVD). Experiments show that the method is successfully applied to non-rigid multi-sensor ensembles and overall yields better registration results than methods that register only two images at a time.

Registering a multisensor ensemble of images.

Many registration scenarios involve aligning more than just two images. These image sets-called ensembles-are conventionally registered by choosing one image as a template, and every other image is registered to it. This pairwise approach is problematic because results depend on which image is chosen as the template. The issue is particularly acute for multisensor ensembles because different sensors create images with different features. Also, pairwise methods use only a fraction of the available data at a time. In this paper, we propose a maximum-likelihood clustering method that registers all the images in a multisensor ensemble simultaneously. Experiments involving rigid-body and affine transformations show that the clustering method is more robust and accurate than competing pairwise registration methods. Moreover, the clustering results can be used to form a rudimentary segmentation of the image ensemble.

Toward a flexible and portable CT scanner.

The very hot and power-hungry x-ray filaments in today's computed tomography (CT) scanners constrain their design to be big and stationary. What if we built a CT scanner that could be deployed at the scene of a car accident to acquire tomographic images before moving the victim? Recent developments in nanotechnology have shown that carbon nanotubes can produce x-rays at room temperature, and with relatively low power needs. We propose a design for a portable and flexible CT scanner made up of an addressable array of tiny x-ray emitters and detectors. In this paper, we outline a basic design, propose a strategy for reconstruction, and demonstrate the concept using a software simulation of the scanner. We also raise a number of issues that still need to be overcome to build such a scanner.

Multimodal image registration using floating regressors in the joint intensity scatter plot.

This paper presents a new approach for multimodal medical image registration and compares it to normalized mutual information (NMI) and the correlation ratio (CR). Like NMI and CR, the new method's measure of registration quality is based on the distribution of points in the joint intensity scatter plot (JISP); compact clusters indicate good registration. This method iteratively fits the JISP clusters with regressors (in the form of points and line segments), and uses those regressors to efficiently compute the next motion increment. The result is a striking, dynamic process in which the regressors float around the JISP, tracking groups of points as they contract into tight clusters. One of the method's strengths is that it is intuitive and customizable, offering a multitude of ways to incorporate prior knowledge to guide the registration process. Moreover, the method is adaptive, and can adjust itself to fit data that does not quite match the prior model. Finally, the method is efficiently expandable to higher-dimensional scatter plots, avoiding the "curse of dimensionality" inherent in histogram-based registration methods such as MI and NMI. In two sets of experiments, a simple implementation of the new registration framework is shown to be comparable to (if not superior to) state-of-the-art implementations of NMI and CR in both accuracy and convergence robustness.

Efficient least squares multimodal registration with a globally exhaustive alignment search.

There are many image registration situations in which the initial misalignment of the two images is large. These registration problems, often involving comparison of the two images only within a region of interest (ROI), are difficult to solve. Most intensity-based registration methods perform local optimization of their cost function and often miss the global optimum when the initial misregistration is large. The registration of multimodal images makes the problem even more difficult since it limits the choice of available cost functions. We have developed an efficient method, capable of multimodal rigid-body registration within an ROI, that performs an exhaustive search over all integer translations, and a local search over rotations. The method uses the fast Fourier transform to efficiently compute the sum of squared differences cost function for all possible integer pixel shifts, and for each shift models the relationship between the intensities of the two images using linear regression. Test cases involving medical imaging, remote sensing and forensic science applications show that the method consistently brings the two images into close registration so that a local optimization method should have no trouble fine-tuning the solution.

Simultaneous registration and activation detection for fMRI.

Registration using the least-squares cost function is sensitive to the intensity fluctuations caused by the blood oxygen level dependent (BOLD) signal in functional MRI (fMRI) experiments, resulting in stimulus-correlated motion errors. These errors are severe enough to cause false-positive clusters in the activation maps of datasets acquired from 3T scanners. This paper presents a new approach to resolving the coupling between registration and activation. Instead of treating the two problems as individual steps in a sequence, they are combined into a single least-squares problem and are solved simultaneously. Robustness tests on a variety of simulated three-dimensional EPI datasets show that the stimulus-correlated motion errors are removed, resulting in a substantial decrease in false-positive and false-negative activation rates. The new method is also shown to decorrelate the motion estimates from the stimulus by testing it on different in vivo fMRI datasets acquired from two different 3T scanners.