Tracking Human activity at home plays a growing factor in fields of security, and of bio-medicine. Microsoft Kinect is a non-wearable sensor that aggregate depth images with traditional optical video frames to estimate individuals' joints' location for kinematic analysis. When the subject of interest is out of Kinect coverage, or not in line of sight, the joints' estimations are distorted, which reduce the estimation accuracy, and can lead, in a scenario of multiple subjects, to erroneous estimations' assignment. In this work we derive features from Kinect joints and form a Kinect Signature (KS). This signature is used to identify different patients, differentiate them from others, exclude artifacts and derive the tracking quality. The suggested technology has the potential to assess human kinematics at home, reduce the cost of the patient traveling to the hospital, and improve the medical treatment follow-up.

When the signal-to-noise ratio (SNR) falls below a certain level, the error of the time-of-flight (ToF) maximum likelihood estimator abruptly increases due to the well-known threshold effect. Nevertheless, operating near and below the threshold SNR value might be necessary for many remote sensing applications due to power-related constraints. These constraints may include a limit on the maximum power of a single source pulse or a limit on the total power used by multiple signals transmitted during a single measurement. For narrow-band signals, the threshold effect emerges mostly due to outliers induced by local maxima of the autocorrelation function of a source signal. Following the previously explored path of biosonar-inspired echo processing, this paper introduces a new method for ToF estimation in the presence of outliers. The proposed method employs a bank of phase-shifted unmatched filters for generating multiple biased but only partially correlated estimators (multiple experts). Using machine learning techniques, the information from the multiple experts is combined together for improving the near-the-threshold ToF estimation from a single echo.

The mean squared error of the classical maximum likelihood time-of-flight (ToF) estimator increases dramatically when the signal-to-noise ratio falls below a certain threshold. For narrow-band signals, this well-known threshold effect occurs largely due to the biased outliers which are induced by the local maxima of the source signal autocorrelation function. In our previous work (Part I), we have described a machine learning biosonar-inspired method for reducing the effect of the outlier bias on the accuracy of a single-echo estimate. In this paper, we extend this approach by introducing a method for combining multiple echo signals into a robust ToF estimator which is resilient to outliers. The individual bias-corrected estimates are combined together using the optimal weighted averaging (OWA) scheme which takes into account the uncertainties associated with inlier and outlier measurements. As in the single-echo case, a weak classifier and a bank of phase-shifted unmatched filters are used for estimating the probabilities of appearance of a specific outlier class in a single-echo estimate. Combined with the previously introduced single-echo bias-reduction method, the OWA scheme results in significant improvement in the ToF estimation accuracy.

Current computational approach to incremental learning requires a constant stream of labelled data to cope with gradual environmental changes known as concept drift. This paper examines a case where labelled data are unavailable. Inspired by the performance of the human visual system, capable of adjusting its concepts using unlabelled stimuli, we introduce a variant to an unsupervised competitive learning algorithm known as the Leader Follower (LF). This variant can adjust pre-learned concepts to environmental changes using unlabelled data samples. We motivate the needed change in the existing LF algorithm and compare between two variants to enable the accumulation of environmental changes when facing unbalanced sample ratio.

Resolution of different types of loops in handwritten script presents a difficult task and is an important step in many classic word recognition systems, writer modeling, and signature verification. When processing a handwritten script, a great deal of ambiguity occurs when strokes overlap, merge, or intersect. This paper presents a novel loop modeling and contour-based handwriting analysis that improves loop investigation. We show excellent results on various loop resolution scenarios, including axial loop understanding and collapsed loop recovery. We demonstrate our approach for loop investigation on several realistic data sets of static binary images and compare with the ground truth of the genuine online signal.

Diffusion Tensor Magnetic Resonance Imaging (DT-MRI) is a non inva- sive method for brain neuronal fibers delineation. Here we show a mod- ification for DT-MRI that allows delineation of neuronal fibers which are infiltrated by edema. We use the Muliple Tensor Variational (MTV) framework which replaces the diffusion model of DT-MRI with a mul- tiple component model and fits it to the signal attenuation with a vari- ational regularization mechanism. In order to reduce free water con- tamination we estimate the free water compartment volume fraction in each voxel, remove it, and then calculate the anisotropy of the remaining compartment. The variational framework was applied on data collected with conventional clinical parameters, containing only six diffusion di- rections. By using the variational framework we were able to overcome the highly ill posed fitting. The results show that we were able to find fibers that were not found by DT-MRI.

In this paper, we present a novel method to extract stroke order independent information from online data. This information, which we term pseudo-online, conveys relevant information on the offline representation of the word. Based on this information, a combination of classification decisions from online and pseudo-online cursive word recognizers is performed to improve the recognition of online cursive words. One of the most valuable aspects of this approach with respect to similar methods that combine online and offline classifiers for word recognition is that the pseudo-online representation is similar to the online signal and, hence, word recognition is based on a single engine. Results demonstrate that the pseudo-online representation is useful as the combination of classifiers perform better than those based solely on pure online information.

An assessment of the performance of a hybrid network with different model selection criteria is considered. These criteria are used in an automatic model selection algorithm for constructing a hybrid network of radial and perceptron hidden units for regression. A forward step builds the full hybrid network; a model selection criterion is used for controlling the network-size and another criterion is used for choosing the appropriate hidden unit for different regions of input space. This is followed by a conservative pruning step using Likelihood Ratio Test, Bayesian or Minimum Description Length, which leads to robust estimators with low variance. The result is a small architecture that performs well on difficult, realistic, benchmark data-sets of high dimensionality and small training size. Best results are obtained by using the Bayesian approach for the model selection.

Using the theory of optimal receivers, the range accuracy of echolocating systems can be expressed as a function of pulse bandwidth and SNR through the well-known Woodward equation. That equation, however, was developed in the limit of very high SNRs and assumes that the correct peak of the cross-correlation function is known a priori. We show that for increasing levels of noise, the accuracy of the cross-correlation receiver undergoes a sharp transition from the Woodward equation for a coherent receiver to a modified Woodward equation for a semicoherent receiver. Since this transition appears at different SNRs for pulses with different center frequencies and bandwidths, it is possible to choose the optimal pulse for any given SNR. We show that the adaptive method we propose outperforms the classical cross-correlation receiver for low SNRs. The same ideas can be applied to the case of a fixed broadband signal, by performing the cross correlation at the receiver end separately in a set of frequency bands with the appropriate center frequencies and bandwidths.

Westudy the selectivity properties of neurons based on BCM and kurtosis energy functions in a general case of noisy high-dimensional input space. The proposed approach, which is used for characterization of the stable states, can be generalized to a whole class of energy functions. We characterize the critical noise levels beyond which the selectivity is destroyed. We also perform a quantitative analysis of such transitions, which shows interesting dependency on data set size. We observe that the robustness to noise of the BCM neuron (Bienenstock, Cooper, & Munro, 1982; Intrator & Cooper, 1992) increases as a function of dimensionality. We explicitly compute the separability limit of BCM and kurtosis learning rules in the case of a bimodal input distribution. Numerical simulations show a stronger robustness of the BCM rule for practical data set size when compared with kurtosis.

To find out how the representations of structured visual objects depend on the co-occurrence statistics of their constituents, we exposed subjects to a set of composite images with tight control exerted over (1) the condi- tional probabilities of the constituent fragments, and (2) the value of Bar- low’s criterion of “suspicious coincidence” (the ratio of joint probability to the product of marginals). We then compared the part verification re- sponse times for various probe/target combinations before and after the exposure. For composite probes, the speedup was much larger for tar- gets that contained pairs of fragments perfectly predictive of each other, compared to those that did not. This effect was modulated by the sig- nificance of their co-occurrence as estimated by Barlow’s criterion. For lone-fragment probes, the speedup in all conditions was generally lower than for composites. These results shed light on the brain’s strategies for unsupervised acquisition of structural information in vision.

We introduce a hybrid recognition/reconstruction architecture that is suitable for recognition of images degraded by various forms of blur. This architecture includes an ensemble of feedforward networks each of which is constrained to reconstruct the inputs in addition to performing classification. The strength of the constraints is controlled by a regularization parameter. Networks are trained on original as well as Gaussian-blurred images, so as to achieve higher robustness to different blur operators. Face recognition is used to demonstrate the proposed method and results are compared to those of classical unconstrained feedforward architectures. In addition, the effect of state-of-the-art restoration methods is demonstrated and it is shown that image restoration with the proposed hybrid architecture leads to the best and most robust results under various forms of blur.

We describe a unified framework for the understanding of struc(cid:173) ture representation in primate vision. A model derived from this framework is shown to be effectively systematic in that it has the ability to interpret and associate together objects that are related through a rearrangement of common "middle-scale" parts, repre(cid:173) sented as image fragments. The model addresses the same concerns as previous work on compositional representation through the use of what+where receptive fields and attentional gain modulation. It does not require prior exposure to the individual parts, and avoids the need for abstract symbolic binding. 1 The problem of structure representation The focus of theoretical discussion in visual object processing has recently started to shift from problems of recognition and categorization to the representation of object structure. Although view- or appearance-based solutions for these problems proved effective on a variety of object classes [1], the "holistic" nature of this approach - the lack of explicit representation of relational structure - limits its appeal as a general framework for visual representation [2]. The main challenges in the processing of structure are productivity and system(cid:173) aticity, two traits commonly attributed to human cognition. A visual system is productive if it is open-ended, that is, if it can deal effectively with a potentially infinite set of objects. A visual representation is systematic if a well-defined change in the spatial configuration of the object (e.g., swapping top and bottom parts) causes a principled change in the representation (e.g., the interchange of the rep(cid:173) resentations of top and bottom parts [3, 2]). A solution commonly offered to the twin problems of productivity and systematicity is compositional representation, in which symbols standing for generic parts drawn from a small repertoire are bound together by categorical symbolically coded relations [4]. 2 The Chorus of Fragments In visual representation, the need for symbolic binding may be alleviated by us(cid:173) ing location in the visual field in lieu of the abstract frame that encodes object structure. Intuitively, the constituents of the object are then bound to each other by virtue of residing in their proper places in the visual field; this can be thought of as a pegboard, whose spatial structure supports the arrangement of parts sus(cid:173) pended from its pegs. This scheme exhibits shallow compositionality, which can be enhanced by allowing the "pegboard" mechanism to operate at different spatial scales, yielding effective systematicity across levels of resolution. Coarse coding the constituents (e.g., representing each object fragment in terms of its similarities to some basis shapes) will render the scheme productive. We call this approach to the representation of structure the Chorus of Fragments (CoF; [5]). 2.1 Neurobiological building blocks What+ Where cells. The representation of spatially anchored object fragments pos(cid:173) tulated by the CoF model can be supported by what+where neurons, each tuned both to a certain shape class and to a certain range of locations in the visual field. Such cells have been found in the monkey in areas V 4 and posterior IT [6], and in the prefrontal cortex [7]. Attentional gain fields. To decouple the representation of object structure from its location in the visual field, one needs a version of the what+where mechanism in which the response of the cell depends not merely on the location of the stimulus with respect to fixation (as in classical receptive fields), but also on its location with respect to the focus of attention. Indeed, modulatory effects of object-centered attention on classical RF structure (gain fields) have been found in area V 4 [8]. 2.2 Implemented model Our implementation of the CoF model involves what+where cells with attention(cid:173) modulated gain fields, and is aimed at productive and systematic treatment of composite shapes in object-centered coordinates. It operates directly on gray-level images, pre-processed by a model of the primary visual cortex [9], with complex(cid:173) cell responses modified to use the MAX operation suggested in [10]. In the model, one what+where unit is assigned to the top and one to the bottom fragment of the visual field, each extracted by an appropriately configured Gaussian gain profile (Figure 2, left). The units are trained (1) to discriminate among five objects, (2) to tolerate translation within the hemifield, and (3) to provide an estimate of the reliability of its output, through an autoassociation mechanism attempting to reconstruct the stimulus image [11, 12]. Within each hemifield, the five outputs of a unit can provide a coarse coding of novel objects belonging to the familiar category, in a manner useful for translation-tolerant recognition [13]. The reliability estimate carries information about category, allowing outputs for objects from other categories to be squelched. Most importantly, due to the spatial localization of the unit's receptive field, the system can distinguish between different configurations of the same shapes, while noting the fragment-wise similarities. We assume that during learning the system performs multiple fixations of the target object, effectively providing the what+where units with a basis for spanning the

Skew detection via principal components is proposed as an effective method for images which contain other parts than text. It is shown that the negative of the image leads to much more robust results, and that the computation time involved is still practical.

We present a new supervised learning procedure for ensemble machines, in which outputs of predictors, trained on different distributions, are combined by a dynamic classifier combination model. This procedure may be viewed as either a version of mixture of experts (Jacobs, Jordan, Nowlan, & Hintnon, 1991), applied to classification, or a variant of the boosting algorithm (Schapire, 1990). As a variant of the mixture of experts, it can be made appropriate for general classification and regression problems by initializing the partition of the data set to different experts in a boostlike manner. If viewed as a variant of the boosting algorithm, its main gain is the use of a dynamic combination model for the outputs of the networks. Results are demonstrated on a synthetic example and a digit recognition task from the NIST database and compared with classifical ensemble approaches.

There is interest in extending the boosting algorithm (Schapire, 1990) to fit a wide range of regression problems. The threshold-based boosting algorithm for regression used an analogy between classification errors and big errors in regression. We focus on the practical aspects of this algorithm and compare it to other attempts to extend boosting to regression. The practical capabilities of this model are demonstrated on the laser data from the Santa Fe times-series competition and the Mackey-Glass time series, where the results surpass those of standard ensemble average.

We study several statistically and biologically motivated learning rules using the same visual environment: one made up of natural scenes and the same single-cell neuronal architecture. This allows us to concentrate on the feature extraction and neuronal coding properties of these rules. Included in these rules are kurtosis and skewness maximization, the quadratic form of the Bienenstock-Cooper-Munro (BCM) learning rule, and single-cell independent component analysis. Using a structure removal method, we demonstrate that receptive fields developed using these rules depend on a small portion of the distribution. We find that the quadratic form of the BCM rule behaves in a manner similar to a kurtosis maximization rule when the distribution contains kurtotic directions, although the BCM modification equations are computationally simpler.

Graphical inspection of multimodality is demonstrated using unsupervised lateral-inhibition neural networks. Three projection pursuit indexes are compared on low-dimensional simulated and real-world data: principal components, Legendre polynomial, and projection pursuit network.

We introduce a new approach to the training of classifiers for performance on multiple tasks. The proposed hybrid training method leads to improved generalization via a better low-dimensional representation of the problem space. The quality of the representation is assessed by embedding it in a two-dimensional space using multi-dimensional scaling, allowing a direct visualization of the results. The performance of the approach is demonstrated on a highly non-linear image classification task.

Bootstrap samples with noise are shown to be an effective smoothness and capacity control technique for training feedforward networks and for other statistical methods such as generalized additive models. It is shown that noisy bootstrap performs best in conjunction with weight-decay regularization and ensemble averaging. The two-spiral problem, a highly non-linear, noise-free data, is used to demonstrate these findings. The combination of noisy bootstrap and ensemble averaging is also shown useful for generalized additive modelling, and is also demonstrated on the well-known Cleveland heart data.

We model a two-eye visual environment composed of natural images and study its effect on single cell synaptic modification. In particular, we study the effect of binocular cortical misalignment on receptive field formation after eye opening. We show that binocular misalignment affects principal component analysis (PCA) and Bienenstock, Cooper, and Munro (BCM) learning in different ways. For the BCM learning rule this misalignment is sufficient to produce varying degrees of ocular dominance, whereas for PCA learning binocular neurons emerge in every case.

Face recognition schemes that are applied directly to gray level pixel images are presented. Two methods for reducing the overfitting-a common problem in high dimensional classification schemes-are presented and the superiority of their combination is demonstrated. The classification scheme is preceded by preprocessing devoted to reducing the viewpoint and scale variability in the data.

We study the effect of removing temporal structure in the prediction error. We observe that networks which are not optimally trained, exhibit strong temporal structure in their prediction error, which can be eliminated using linear regression. This elimination improves performance significantly, but does not lead to the best performance which is achieved by training networks until they do not exhibit any such temporal structure. The improvement in performance of ensemble net averaging does not affect possible temporal structure of the error, thus averaging can be performed before or after temporal structure removal. We demonstrate these findings on the sunspot data set.

We present a novel classification and regression method that combines exploratory projection pursuit (unsupervised training) with projection pursuit regression (supervised training), to yield a new family of cost/complexity penalty terms. Some improved generalization properties are demonstrated on real-world problems.

We propose an object recognition scheme based on a method for feature extraction from gray level images that corresponds to recent statistical theory, called projection pursuit, and is derived from a biologically motivated feature extracting neuron. To evaluate the performance of this method we use a set of very detailed psychophysical three-dimensional object recognition experiments (Bülthoff and Edelman 1992).

A novel unsupervised neural network for dimensionality reduction that seeks directions emphasizing multimodality is presented, and its connection to exploratory projection pursuit methods is discussed. This leads to a new statistical insight into the synaptic modification equations governing learning in Bienenstock, Cooper, and Munro (BCM) neurons (1982). The importance of a dimensionality reduction principle based solely on distinguishing features is demonstrated using a phoneme recognition experiment. The extracted features are compared with features extracted using a backpropagation network.