Neural associative memories (NAM) are perceptron-like single-layer networks with fast synaptic learning typically storing discrete associations between pairs of neural activity patterns. Gripon and Berrou ( 2011 ) investigated NAM employing block coding, a particular sparse coding method, and reported a significant increase in storage capacity. Here we verify and extend their results for both heteroassociative and recurrent autoassociative networks. For this we provide a new analysis of iterative retrieval in finite autoassociative and heteroassociative networks that allows estimating storage capacity for random and block patterns. Furthermore, we have implemented various retrieval algorithms for block coding and compared them in simulations to our theoretical results and previous simulation data. In good agreement of theory and experiments, we find that finite networks employing block coding can store significantly more memory patterns. However, due to the reduced information per block pattern, it is not possible to significantly increase stored information per synapse. Asymptotically, the information retrieval capacity converges to the known limits [Formula: see text] and [Formula: see text] also for block coding. We have also implemented very large recurrent networks up to [Formula: see text] neurons, showing that maximal capacity [Formula: see text] bit per synapse occurs for finite networks having a size [Formula: see text] similar to cortical macrocolumns.

In this paper, a robust system for view-independent action unit intensity estimation is presented. Based on the theory of sparse coding, region-specific dictionaries are trained to approximate the characteristic of the individual action units. The system incorporates landmark detection, face alignment and contrast normalization to handle a large variety of different scenes. Coupled with head pose estimation, an ensemble of large margin classifiers is used to detect the individual action units. The experimental validation shows that our system is robust against pose variations and able to outperform the challenge baseline by more than 35%.

A new algorithm named EXPected Similarity Estimation (EXPOSE) was recently proposed to solve the problem of large-scale anomaly detection. It is a non-parametric and distribution free kernel method based on the Hilbert space embedding of probability measures. Given a dataset of n samples, EXPOSE takes 𝒪(n) time to build a model and 𝒪(1) time per prediction.

Learning from demonstration (LfD) aims at robots learning skills from human-demonstrated tasks. Robots should be able to learn at all levels of abstraction. Unlike at the level of motor primitives, high-level LfD requires symbolic representations. It thus faces the classical problem of symbol grounding. Furthermore, it requires the robot to interpret human-demonstrated actions at a higher, conceptual abstraction level. We present a method, that enables a robot to recognize human-demonstrated pick-and-place task goals on an object-relational abstraction layer. The robot can reproduce the task goals in new situations using a symbolic planner. We show that in a robotic context conceptual spaces can serve as a mean for symbol grounding at an object-relational level as well as for the recognition of conceptual similarities in effects of human-demonstrated actions. The method is evaluated in experiments on a real robot.

We propose a new algorithm named EXPected Similarity Estimation (EXPoSE) to approach the problem of anomaly detection (also known as one-class learning or outlier detection) which is based on the similarity between data points and the distribution of non-anomalous data. We formulate the problem as an inner product in a reproducing kernel Hilbert space to which we present approximations that allow its application to very large-scale datasets. More precisely, given a dataset with n instances, our proposed method requires O(n) training time and O(1) to make a prediction while spending only O(1) memory to store the learned model. Despite its abstract derivation our algorithm is simple and parameter free. We show on seven real datasets that our approach can compete with state of the art algorithms for anomaly detection.

Recent trends in human–computer interaction (HCI) show a development towards cognitive technical systems (CTS) to provide natural and efficient operating principles. To do so, a CTS has to rely on data from multiple sensors which must be processed and combined by fusion algorithms. Furthermore, additional sources of knowledge have to be integrated, to put the observations made into the correct context. Research in this field often focuses on optimizing the performance of the individual algorithms, rather than reflecting the requirements of CTS. This paper presents the information fusion principles in CTS architectures we developed for Companion Technologies. Combination of information generally goes along with the level of abstractness, time granularity and robustness, such that large CTS architectures must perform fusion gradually on different levels — starting from sensor-based recognitions to highly abstract logical inferences. In our CTS application we sectioned information fusion approaches into three categories: perception-level fusion, knowledge-based fusion and application-level fusion. For each category, we introduce examples of characteristic algorithms. In addition, we provide a detailed protocol on the implementation performed in order to study the interplay of the developed algorithms.

Only few cognitive architectures have been proposed that cover the complete range from recognizers working on the direct sensor input, to logical inference mechanisms of classical artificial intelligence (AI). Logical systems operate on abstract predicates, which are often related to an action-like state transition, especially when compared to the classes recognized by pattern recognition approaches. On the other hand, pattern recognition is often limited to static patterns, and temporal and multi-modal aspects of a class are often not regarded, e.g. by testing only on pre-segmented data. Recent trends in AI aim at developing applications and methods that are motivated by data-driven real world scenarios, while the field of pattern recognition attempts to push forward the boundary of pattern complexity. We propose a new generic architecture to close the gap between AI and pattern recognition approaches. In order to detect abstract complex patterns, we process sequential data in layers. On each layer, a set of elementary classes is recognized and the outcome of the classification is passed to the successive layer such that the time granularity increases. Layers can combine modalities, additional symbolic information or make use of reasoning algorithms. We evaluated our approach in an on-line scenario of activity recognition using three layers. The obtained results show that the combination of concepts from pattern recognition and high-level symbolic information leads to a prosperous and powerful symbiosis.

Human action recognition in video streams is a fast developing field in pattern recognition and machine learning. Local image representations, e.g. space-time interest points [1], have proven to be the current most reliable choice of feature in sequences in which the region of interest is difficult to determine [2]. However, the question how to deal with more severe occlusions has large been ignored [2]. This work proposes a new approach which directly addresses heavy occlusions by modeling the skeleton-based features using a probability density functions (PDF) defined over graphs. We integrated the proposed density into an hidden Markov model (HMM) to model sequences of graphs of arbitrary sizes, i.e. occlusions setting may change over time. The approach is evaluated using a dataset embracing three action classes, studying six different types of occlusions (involving the removal of subgraphs from the graphical representation of action sequence). The presented study shows clearly that actions from even heavily occluded sequences can be reliably recognized.

Learning Convolutional Neural Networks (CNN) is commonly carried out by plain supervised gradient descent. With sufficient training data, this leads to very competitive results for visual recognition tasks when starting from a random initialization. When the amount of labeled data is limited, CNNs reveal their strong dependence on large amounts of training data. However, recent results have shown that a well chosen optimization starting point can be beneficial for convergence to a good generalizing minimum. This starting point was mostly found using unsupervised feature learning techniques such as sparse coding or transfer learning from related recognition tasks. In this work, we compare these two approaches against a simple patch based initialization scheme and a random initialization of the weights. We show that pre-training helps to train CNNs from few samples and that the correct choice of the initialization scheme can push the network's performance by up to 41% compared to random initialization.

The audio/visual emotion challenge (AVEC) resembles a benchmarking data collection in order to evaluate and develop techniques for the recognition of affective states. In our work, we present a Markov fusion network (MFN) for the combination of different individual classifiers, that is derived from the well-known Markov random fields (MRF). It is capable to restore missing values from a sequence of decisions and can integrate multiple channels and weights them dynamically using confidences. The approach shows promising challenge results compared to the baseline.

It is essential for the advancement of human-centered multimodal interfaces to be able to infer the current user's state or communication state. In order to enable a system to do that, the recognition and interpretation of multimodal social signals (i.e., paralinguistic and nonverbal behavior) in real-time applications is required. Since we believe that laughs are one of the most important and widely understood social nonverbal signals indicating affect and discourse quality, we focus in this work on the detection of laughter in natural multiparty discourses. The conversations are recorded in a natural environment without any specific constraint on the discourses using unobtrusive recording devices. This setup ensures natural and unbiased behavior, which is one of the main foci of this work. To compare results of methods, namely Gaussian Mixture Model (GMM) supervectors as input to a Support Vector Machine (SVM), so-called Echo State Networks (ESN), and a Hidden Markov Model (HMM) approach, are utilized in online and offline detection experiments. The SVM approach proves very accurate in the offline classification task, but is outperformed by the ESN and HMM approach in the online detection (F1scores: GMM SVM 0.45, ESN 0.63, HMM 0.72). Further, we were able to utilize the proposed HMM approach in a cross-corpus experiment without any retraining with respectable generalization capability (F1score: 0.49). The results and possible reasons for these outcomes are shown and discussed in the article. The proposed methods may be directly utilized in practical tasks such as the labeling or the online detection of laughter in conversational data and affect-aware applications.

Non-negative Matrix Factorization is a technique for decomposing large data sets into bases and code words, where all entries of the occurring matrices are non-negative. A recently proposed technique also incorporates sparseness constraints, in such a way that the amount of nonzero entries in both bases and code words becomes controllable. This paper extends the Non-negative Matrix Factorization with Sparseness Constraints. First, a modification of the optimization criteria ensures fast inference of the code words. Thus, the approach is real-time capable for use in time critical applications. Second, in case a teacher signal is associated with the samples, it is considered in order to ensure that inferred code words of different classes can be well distinguished. Thus, the derived bases generate discriminative code words, which is a crucial prerequisite for training powerful classifiers. Experiments on natural image patches show, similar to recent results in the field of sparse coding algorithms, that Gabor-like filters are minimizing the reconstruction error while retaining inference capabilities. However, applying the approach with incorporation of the teacher signal to handwritten digits yields morphologically completely different bases, while achieving superior classification results.

This contribution presents an efficient and useful way to readapt a cascaded classifier. We introduce Package Boosting which combines the advantages of Real Adaboost and Online Boosting for the realization of the strong learners in each cascade layer. We also examine the conditions which need to be fulfilled by a cascade in order to meet the requirements of an online algorithm and present the evaluation results of the system.

In the present paper an application of a novel neural network architecture called Echo State Network (ESN) within the frame of a reinforcement learning scheme named Adaptive Critic Design (ACD) is proposed. Our aim is to investigate the possibility for on-line training of adaptive critic using the ESN architecture. In particular the application of this approach to mobile robot control is presented. Our preliminary results are encouraging and demonstrate that ESNs are good candidates for the on-line application of an ACD optimization approach due to their specific structure and fast training algorithm.

Neural associative networks with plastic synapses have been proposed as computational models of brain functions and also for applications such as pattern recognition and information retrieval. To guide biological models and optimize technical applications, several definitions of memory capacity have been used to measure the efficiency of associative memory. Here we explain why the currently used performance measures bias the comparison between models and cannot serve as a theoretical benchmark. We introduce fair measures for information-theoretic capacity in associative memory that also provide a theoretical benchmark. In neural networks, two types of manipulating synapses can be discerned: synaptic plasticity, the change in strength of existing synapses, and structural plasticity, the creation and pruning of synapses. One of the new types of memory capacity we introduce permits quantifying how structural plasticity can increase the network efficiency by compressing the network structure, for example, by pruning unused synapses. Specifically, we analyze operating regimes in the Willshaw model in which structural plasticity can compress the network structure and push performance to the theoretical benchmark. The amount C of information stored in each synapse can scale with the logarithm of the network size rather than being constant, as in classical Willshaw and Hopfield nets (< or = ln 2 approximately 0.7). Further, the review contains novel technical material: a capacity analysis of the Willshaw model that rigorously controls for the level of retrieval quality, an analysis for memories with a nonconstant number of active units (where C < or = 1/e ln 2 approximately 0.53), and the analysis of the computational complexity of associative memories with and without network compression.

In this paper we investigate how neural fields can generate complex behavior in mobile robot navigation. A case study is presented in which a mobile robot endowed with sonars and a laser scanner performs the behaviorstarget-acquisition,obstacle-avoidance, andsubtarget-selection. The design of these behaviors as well as their coordination are achieved through correct choice of stimulus parameters on the field. We describe the approach in theoretical terms, supported by experimental results.

We have implemented a system that can understand spoken command sentences like "Bot lift green apple" using hidden Markov models (HMMs) and neural associative memories. After speaking a command sentence into a microphone, the system processes it in three stages: As first step, the auditory input is transformed into a convenient subsymbolic representation (diphones or triphones) using HMMs. The second step retrieves a symbolic representation (words) from the subsymbolic representation using a network of neural associative memories. Finally, in step three a semantic representation is obtained using neural associative memories. Furthermore, the system can learn new object words during performance.

In most behavior-based approaches, implementing a broad set of different behavioral skills and coordinating them to achieve coherent complex behavior is an error-prone and very tedious task. Concepts for organizing reactive behavior in a hierarchical manner are rarely found in behavior-based approaches, and there is no widely accepted approach for creating such behavior hierarchies. Most applications of behavior-based concepts use only few behaviors and do not seem to scale well. Reuse of behaviors for different application scenarios or even on different robots is very rare, and the integration of behavior-based approaches with planning is unsolved. This paper discusses the design, implementation, and performance of a behavior framework that addresses some of these issues within the context of behavior-based and hybrid robot control architectures. The approach presents a step towards more systematic software engineering of behavior-based robot systems.

We present different training algorithms for radial basis function (RBF) networks. The behaviour of RBF classifiers in three different pattern recognition applications is presented: the classification of 3-D visual objects, high-resolution electrocardiograms and handwritten digits.

A novel method, referred to as group vector quantization (GVQ), is proposed to train VQ codebooks for closed-set speaker identification. In GVQ training, speaker codebooks are optimized for vector groups rather than for individual vectors. An evaluation experiment has been conducted to compare the codebooks trained by the Linde-Buzo-Grey (LBG), the learning vector quantization (LVQ), and the GVQ algorithms. It is shown that the frame scores from the GVQ trained codebooks are less correlated, therefore, the sentence level speaker identification rate increases more quickly with the length of test sentences.

We present rules for the unsupervised learning of coincidence between excitatory postsynaptic potentials (EPSPs) by the adjustment of postsynaptic delays between the transmitter binding and the opening of ion channels. Starting from a gradient descent scheme, we develop a robust and more biological threshold rule by which EPSPs from different synapses can be gradually pulled into coincidence. The synaptic delay changes are determined from the summed potential--at the site where the coincidence is to be established--and from postulated synaptic learning functions that accompany the individual EPSPs. According to our scheme, templates for the detection of spatiotemporal patterns of synaptic activation can be learned, which is demonstrated by computer simulation. Finally, we discuss possible relations to biological mechanisms.

VQ-based speaker recognition has proven to be a successful method. Usually, a codebook is trained to minimize the quantization error for the data from an individual speaker. The codebooks trained based on this criterion have weak discriminative power when used as a classifier. The LVQ algorithm can be used to globally train the VQ-based classifier. However, the correlation between the feature vectors is not taken into consideration, in consequence, a high classification rate for feature vectors does not lead to a high classification rate for the test sentences. A heuristic training procedure is proposed to retrain the codebooks so that they give a lower classification error rate for randomly selected vector-groups. Evaluation experiments demonstrated that the codebooks trained with this method provide much higher recognition rates than that trained with the LBG algorithm alone, and often they can outperform the more powerful Gaussian mixture speaker models.

For technical speech recognition systems as well as for humans it has been shown that the combination of acoustic and optic information can enhance speech recognition performance. But it still remains an open question, at which stage of processing the two information channels should be combined. We systematically investigate this problem by means of a neural speech recognition system applied to monosyllabic words. Different fusion architectures of multilayer perceptrons are compared both for noiseless and noisy acoustic data. Furthermore, different modularized neural architectures are examined for the acoustic channel alone. The results corroborate the idea of separate processing of the two channels until the final stage of classification.

In speech recognition based on LPC analysis the prediction residues are usually ignored, only the LPC-derived cepstral coefficients (LPCC) are used to compose feature vectors. In this study, a number of parameters (called the residual cepstrum or RCEP) were calculated from these residues and their effectiveness for speech recognition was evaluated. It was shown that the RCEP do contain useful information, in particular, they are complementary to the LPCC. In an evaluation experiment, if the LPCC were used jointly with a few RCEP coefficients, the recognition rate of the English E-set letters was improved from 54% to 67% and from 69% to 71% by the use of HMMs based recognizer and the DTW based recognizer, respectively. In addition, Mel-scaled FFT based cepstrum (MFCC) was found to be superior to LPCC.

A large number of parameters, including pitch, LPCC, /spl Delta/LPCC, PARCOR, MFCC, /spl Delta/MFCC, and residual cepstrum (RCEP) were extracted from speech signals and their effectiveness for text-independent speaker identification was evaluated. In addition, the usefulness of two signal processing techniques, preemphasis and cepstral weighting, was also studied. The VQ-based speaker recognition method with codebooks fine-tuned by LVQ algorithm was used. It was shown that both LPCC and MFCC are effective representations, for smaller number of parameters, LPCC representation performs better but is surpassed by MFCC if the analysis order is larger. Pitch is an independent parameter so that it can be used jointly with other spectral features. In an evaluation experiment, the correct identification rate for 112 male speakers with test utterances of less than one second reached 98.2%.

A simple relation between the storage capacity A for autoassociation and H for heteroassociation with a local learning rule is demonstrated: H = 2A. Both values are bounded by local learning bounds: A ≤ L A and H ≤ L H. L H = 2L A is evaluated numerically.