To adapt to specific places and people, robots must recognize objects, which are typically taught by users-a tedious process.Inspired by anecdotes of positive teaching experiences shared by educators, sports coaches, and animal trainers, we developed seven experience-driven ways to make teaching a robot more engaging.For example, one interaction involved the robot prompting users to tell personal stories about the objects.A video vignette study (N=184) showed that experience-driven teaching was perceived as more positive than the current technology-driven teaching.Participants reported feeling more competent, connected to the robot, and valued.Additionally, the robot was perceived as more extroverted, open, agreeable, and conscientious.Overall, the experience-driven design of teaching interactions enhanced engagement and persistence by fostering reciprocal exchange and mutual understanding.In addition, the study lends support to an anecdotal approach to designing positive experiences through technology.

AURORAis a software platform, that facilitates scalable deployment of robotic simulations over the web for the Human-Robot Interaction (HRI) community. As robotics is becoming increasingly important in various disciplines, there is a growing need for accessible and scalable research methods. Traditional experiments often require expensive hardware and in-person participation, limiting accessibility and participant diversity. Our platform allows researchers from different fields to easily provide HRI experiences by deploying online studies with robotic simulations paired with customizable surveys, allowing end users worldwide to interact with these simulations. Our platform is entirely open source and can be hosted locally, providing flexibility and control of the research environment. Since AURORA is implemented with Docker, it is platform-independent. By offering a user-friendly interface that can be deployed and used without extensive technical expertise, our platform reduces costs, increases participant diversity, and improves the reproducibility of research in the HRI community.

Judith Sieker, Simeon Junker, Ronja Utescher, Nazia Attari, Heiko Wersing, Hendrik Buschmeier, Sina Zarrieß. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

This paper introduces the Mental Model Mismatch (MMM) Score, a feedback mechanism designed to align human teaching behavior with robot learning by quantifying mismatches between the human teacher’s mental model and the robot’s learning capabilities. Utilizing a LLM, the system analyzes human teacher intentions in natural language to generate adaptive feedback for the human. A study with 150 participants teaching a virtual robot learner demonstrates that intention-based feedback significantly improves the robots learning outcomes compared to traditional performance-based feedback or no feedback. The findings suggest that this approach improves understanding of the robot’s learning process and reduces misconceptions to enhance human-robot interaction.

How can a robot provide unobtrusive physical support within a group of humans? We present Attentive Support, a novel interaction concept for robots to support a group of humans. It combines scene perception, dialogue acquisition, situation understanding, and behavior generation with the common-sense reasoning capabilities of Large Language Models (LLMs). In addition to following user instructions, Attentive Support is capable of deciding when and how to support the humans, and when to remain silent to not disturb the group. With a diverse set of scenarios, we show and evaluate the robot’s attentive behavior, which supports and helps the humans when required, while not disturbing if no help is needed.

While machine learning (ML) gives rise to astonishing results in automated systems, it is usually at the cost of large data requirements. This makes many successful algorithms from ML unsuitable for human-machine interaction, where the machine must learn from a small number of training samples that can be provided by a user within a reasonable time frame. Fortunately, the user can tailor the training data they create to be as useful as possible, severely limiting its necessary size—as long as they know about the machine’s requirements and limitations. Of course, acquiring this knowledge can in turn be cumbersome and costly. This raises the question of how easy ML algorithms are to interact with. In this work, we address this issue by analyzing the intuitiveness of certain algorithms when they are actively taught by users. After developing a theoretical framework of intuitiveness as a property of algorithms, we introduce an active teaching paradigm involving a prototypical two-dimensional spatial learning task as a method to judge the efficacy of human-machine interactions. Finally, we present and discuss the results of a large-scale user study into the performance and teaching strategies of 800 users interacting with two prominent ML algorithms in our system, providing first evidence for the role of intuition as an important factor impacting human-machine interaction.

Convolutional neural networks have been used to achieve a string of successes during recent years, but their lack of interpretability remains a serious issue. Adversarial examples are designed to deliberately fool neural networks into making any desired incorrect classification, potentially with very high certainty. Several defensive approaches increase robustness against adversarial attacks, demanding attacks of greater magnitude, which lead to visible artifacts. By considering human visual perception, we compose a technique that allows to hide such adversarial attacks in regions of high complexity, such that they are imperceptible even to an astute observer. We carry out a user study on classifying adversarially modified images to validate the perceptual quality of our approach and find significant evidence for its concealment with regards to human visual perception.

Real-time learning from data streams in non-stationary environments gains ever more relevance due to the exponentially increasing amounts of generated data. Recently, the Self-Adjusting Memory (SAM) was proposed, an algorithm able to robustly handle heterogeneous types on the basis of two dedicated memories for the current and former concepts that continuously preserve consistency with explicit filtering. Yet, since the algorithm is restricted to one memory architecture, the variety of possible alternatives is limited by design in favor of an overall model consistency. Moreover, it does not actively detect drift, thus adapting with a relatively high delay in case of abrupt changes. We propose a dynamic ensemble on the basis of the SAM algorithm, which is triggered by both, the inherent passive adaptation of SAM and active drift detection. Further, since SAM is based on the stable k-Nearest-Neighbor algorithm, we investigate multiple approaches to obtain a high diversity in the ensemble, resulting in an effective overall strategy. The increased computational demand is countered on the basis of a parallel implementation. We extensively evaluate the method on numerous benchmarks, where it consistently achieves superior results in comparison to state-of-the-art methods.

Online action classification is an important field of research, enabling the particularly interesting application scenario of controlling wearable devices which actively support the user's motions. The majority of machine learning applications of real-world systems are based on pre-trained average-user models without any personalization. Our long-term goal is to provide a system that adapts to its user's personal behavior patterns on the fly and in real-time. Ideally, we want to initiate a continuous collaboration between the system and the user where both alternatively adjust to each other to maximize the system's utility. Such tasks are not feasible with static models. In this paper, we investigate the potential and benefits of personalized online learning in the task of online action classification. We record motion sequences of different subjects wearing the Xsens bodysuit, which incorporates multiple inertial measuring units, enabling a fine-grained discrimination of motions. On this basis, we first perform a feature selection, showing that only a few sensors are necessary to achieve a high classification performance. Subsequently, we compare the recognition capabilities of offline average user models against personalized models trained in an online way. Our experiments conclude that personalized models require only few data to outperform average user systems and are particularly valuable for applications with limited computational hardware which rely on the raw sensor inputs only.

An increasing number of industrial areas recognize the opportunities of Big Data, requiring highly efficient algorithms which enable real-time processing to reduce the burden of data storage and maintenance. Decision trees are extremely fast, highly accurate and easy to use in practice. Merging multiple decision trees to an ensemble leads to one of the most powerful machine learning methods. The Very Fast Decision Tree is the state-of-the-art incremental decision tree induction algorithm, capable of learning from massive data streams. It is successful due to its theoretical guarantees based on the Hoeffding bound as well as its competitive performance in terms of classification accuracy and time / space efficiency. In this paper, we increase the efficiency even further by replacing its global splitting scheme, which periodically tries to split every nminexamples. Instead, we utilize local statistics to predict the split-time, thus, avoiding unnecessary split-attempts, usually dominating the computational cost. Concretely, we use the class distributions of previous split-attempts to approximate the minimum number of examples until the Hoeffding bound is met. This cautious approach yields by design a low delay and reduces the number of split-attempts at the same time. We extensively evaluate our method using common stream-learning benchmarks also considering non-stationary environments. The experiments confirm a substantially reduced run-time without a loss in classification performance.

Endowing assistant systems with more autonomy establishes the transition from a human-controlled tool towards a self-directed agent capable of own decisions and goals. In this concept paper we suggest to perform the design of such an assistant agent according to principles of cooperativity. We first review definitions of cooperation between animals, humans and machines and then discuss advantages of cooperation also for a human-machine interaction system. We concentrate on the important roles of adaptivity and responsibility within the interaction. We argue that main benefits of a cooperative design are alleviation of typical automation issues like controllability, complacency, trust, and greater flexibility of the combined human-machine system in tasks with high variability.

Data Mining in non-stationary data streams is particularly relevant in the context of the Internet of Things and Big Data. Its challenges arise from fundamentally different drift types violating assumptions of data independence or stationarity. Available methods often struggle with certain forms of drift or require unavailable a priori task knowledge. We propose the Self-Adjusting Memory (SAM) model for the k Nearest Neighbor (kNN) algorithm. SAM-kNN can deal with heterogeneous concept drift, i.e. different drift types and rates. Its basic idea are dedicated models for current and former concepts used according to the demands of the given situation. It can be robustly applied in practice without meta parameter optimization. We conduct an extensive evaluation on various benchmarks, consisting of artificial streams with known drift characteristics and real-world datasets. Highly competitive results throughout all experiments underline the robustness of SAM-kNN as well as its capability to handle heterogeneous concept drift.

Data Mining in non-stationary data streams is gaining more attentionrecently, especially in the context of Internet of Things and Big Data. It is a highly challenging task, since the fundamentally different typesof possibly occurring drift undermine classical assumptions such asi.i.d. data or stationary distributions. Available algorithms are either struggling with certain forms of drift or require a priori knowledge in terms of a task specific setting. We propose the Self Adjusting Memory (SAM) model for the k Nearest Neighbor (kNN) algorithm since kNN constitutes a proven classifier within the streaming setting. SAM-kNN can deal with heterogeneous concept drift, i.e different drift types and rates, using biologically inspiredmemory models and their coordination. It can be easilyapplied in practice since an optimization of the meta parameters is not necessary. The basic idea is to construct dedicated models for thecurrent and former concepts and apply them according tothe demands of the given situation. An extensive evaluation on various benchmarks, consisting of artificial streamswith known drift characteristics as well as real world datasets is conducted. Thereby, we explicitly add new benchmarks enabling a precise performance evaluation on multiple types of drift. The highly competitive results throughout all experiments underline the robustness of SAM-kNN as well as its capabilityto handle heterogeneous concept drift.

One of the main aims of lifelong learning architectures is to efficiently and reliably cope with the stability-plasticity dilemma. A viable solution of this dilemma combines a static offline classifier, which preserves ground knowledge that should be respected during training, with an incremental online learning of new or specific information encountered during use. A feasible realisation has been published lately based on intuitive distance-based classifiers using the concept of metric learning (Fischer et al.: Combining offline and online classifiers for life-long learning (OOL), IJCNN'15). One crucial aspect of such a system is how to combine the offline and online model. A generic approach, taken in OOL, uses a dynamic classifier selection strategy based on confidences of both classifiers. This can cause problems in the case of confidence drift, especially when the validity of the confidence estimation of the static offline classifier changes. This pitfall occurs in the context of metric learning whenever the metric tensor of the online system becomes orthogonal to the metric of the offline system, hence the respective internal data description mismatch. We propose an efficient metric learning strategy which allows an online adaptation of an invalid confidence estimation of the OOL architecture in case of confidence drift.

One of the greatest challenges of life-long learning architectures is how to efficiently and reliably cope with the stability-plasticity dilemma. We propose an extension of a flexible system combining a static offline classifier and an incremental online classifier that is well suited for life-long learning scenarios. The pre-trained offline classifier preserves ground knowledge that should be respected during training, while the online classifier enables learning of new or specific information encountered during use. The combination is realised by a dynamic classifier selection strategy based on confidences of both ingredients. We report exemplary results of this architecture for the case of learning vector quantization (LVQ) for several data sets, thereby including an extensive comparison to alternative state of the art algorithms for incremental learning such as incremental generalised LVQ and the support vector machine.

We present an architecture for incremental online learning in high-dimensional feature spaces and apply it on a mobile robot. The model is based on learning vector quantization, approaching the stability-plasticity problem of incremental learning by adaptive insertions of representative vectors. We employ a cost-function-based learning vector quantization approach and introduce a new insertion strategy optimizing a cost-function based on a subset of samples. We demonstrate this model within a real-time application for a mobile robot scenario, where we perform interactive real-time learning of visual categories.

Facial communicative signals (FCSs) such as head gestures, eye gaze, and facial expressions can provide useful feedback in conversations between people and also in human-robot interaction. This paper presents a pattern recognition approach for the interpretation of FCSs in terms of valence, based on the selection of discriminative subsequences in video data. These subsequences capture important temporal dynamics and are used as prototypical reference subsequences in a classification procedure based on dynamic time warping and feature extraction with active appearance models. Using this valence classification, the robot can discriminate positive from negative interaction situations and react accordingly. The approach is evaluated on a database containing videos of people interacting with a robot by teaching the names of several objects to it. The verbal answer of the robot is expected to elicit the display of spontaneous FCSs by the human tutor, which were classified in this work. The achieved classification accuracies are comparable to the average human recognition performance and outperformed our previous results on this task.

We study the problem of actively searching for an object in a three-dimensional (3-D) environment under the constraint of a maximum search time using a visually guided humanoid robot with 26 degrees of freedom. The inherent intractability of the problem is discussed, and a greedy strategy for selecting the best next viewpoint is employed. We describe a target probability updating scheme approximating the optimal solution to the problem, providing an efficient solution to the selection of the best next viewpoint. We employ a hierarchical recognition architecture, inspired by human vision, that uses contextual cues for attending to the view-tuned units at the proper intrinsic scales and for active control of the robotic platform sensor's coordinate frame, which also gives us control of the extrinsic image scale and achieves the proper sequence of pathognomonic views of the scene. The recognition model makes no particular assumptions on shape properties like texture and is trained by showing the object by hand to the robot. Our results demonstrate the feasibility of using state-of-the-art vision-based systems for efficient and reliable object localization in an indoor 3-D environment.

We introduce a greedy algorithm that works from coarse to fine by iteratively applying localized principal component analysis (PCA). The decision where and when to split or add new components is based on two antagonistic criteria. Firstly, the well known quadratic reconstruction error and secondly a measure for the homogeneity of the distribution. For the latter criterion, which we call “generation error”, we compared two different possible methods to assess if the data samples are distributed homogeneously. The proposed algorithm does not involve a costly multi-objective optimization to find a partition of the inputs. Further, the final number of local PCA units, as well as their individual dimensionality need not to be predefined. We demonstrate that the method can flexibly react to different intrinsic dimensionalities of the data.

We present an exemplar-based learning approach for incremental and life-long learning of visual categories. The basic concept of the proposed learning method is to subdivide the learning process into two phases. In the first phase we utilize supervised learning to generate an appropriate category seed, while in the second phase this seed is used to autonomously bootstrap the visual representation. This second learning phase is especially useful for assistive systems like a mobile robot, because the visual knowledge can be enhanced even if no tutor is present. Although for this autonomous bootstrapping no category labels are provided, we argue that contextual information is beneficial for this process. Finally we investigate the effect of the proposed second learning phase with respect to the overall categorization performance.

In everyday conversation besides speech people also communicate by means of nonverbal cues. Facial expressions are one important cue, as they can provide useful information about the conversation, for instance, whether the interlocutor seems to understand or appears to be puzzled. Similarly, in human-robot interaction facial expressions also give feedback about the interaction situation. We present a Wizard of Oz user study in an object-teaching scenario where subjects showed several objects to a robot and taught the objects' names. Afterward, the robot should term the objects correctly. In a first evaluation, we let other people watch short video sequences of this study. They decided by looking at the face of the human whether the answer of the robot was correct (unproblematic situation) or incorrect (problematic situation). We conducted the experiments under specific conditions by varying the amount of temporal and visual context information and compare the results with related experiments described in the literature.

Object representation in the inferior temporal cortex (IT), an area of visual cortex critical for object recognition in the primate, exhibits two prominent properties: (1) objects are represented by the combined activity of columnar clusters of neurons, with each cluster representing component features or parts of objects, and (2) closely related features are continuously represented along the tangential direction of individual columnar clusters. Here we propose a learning model that reflects these properties of parts-based representation and topographic organization in a unified framework. This model is based on a nonnegative matrix factorization (NMF) basis decomposition method. NMF alone provides a parts-based representation where nonnegative inputs are approximated by additive combinations of nonnegative basis functions. Our proposed model of topographic NMF (TNMF) incorporates neighborhood connections between NMF basis functions arranged on a topographic map and attains the topographic property without losing the parts-based property of the NMF. The TNMF represents an input by multiple activity peaks to describe diverse information, whereas conventional topographic models, such as the self-organizing map (SOM), represent an input by a single activity peak in a topographic map. We demonstrate the parts-based and topographic properties of the TNMF by constructing a hierarchical model for object recognition where the TNMF is at the top tier for learning high-level object features. The TNMF showed better generalization performance over NMF for a data set of continuous view change of an image and more robustly preserving the continuity of the view change in its object representation. Comparison of the outputs of our model with actual neural responses recorded in the IT indicates that the TNMF reconstructs the neuronal responses better than the SOM, giving plausibility to the parts-based learning of the model.

During CAD development and any kind of design optimisation over years a huge amount of geometries accumulate in a design department. To organize and structure these designs with respect to reusability, a hierarchical set of components on different scalings is extracted by the designers. This hierarchy allows to compose designs from several parts and to adapt the composition to the current task. Nevertheless, this hierarchy is imposed by humans and relies on their experiences. In the present paper a computational method is proposed for an unsupervised extraction of design components from a large repository of geometries. Methods known from the field of object and pattern recognition in images are transferred to the 3D design space to detect relevant features of geometries. The non-negative matrix factorization algorithm (NMF) is extended and tuned to the given task for an autonomous detection of design components. The results of the NMF additionally provide an overview on the distribution of these components in the design repository. The extracted components sum up in a parts-based representation which serves as a base for manual or computational design development or optimisation respectively.

We present a biologically motivated architecture for object recognition that is capable of online learning of several objects based on interaction with a human teacher. The system combines biological principles such as appearance-based representation in topographical feature detection hierarchies and context-driven transfer between different levels of object memory. Training can be performed in an unconstrained environment by presenting objects in front of a stereo camera system and labeling them by speech input. The learning is fully online and thus avoids an artificial separation of the interaction into training and test phases. We demonstrate the performance on a challenging ensemble of 50 objects.

Sparse coding is an important approach for the unsupervised learning of sensory features. In this contribution, we present two new methods that extend the traditional sparse coding approach with supervised components. Our goal is to increase the suitability of the learned features for classification tasks while keeping most of their general representation capability. We analyze the effect of the new methods using visualization on artificial data and discuss the results on two object test sets with regard to the properties of the found feature representation.

We present a hybrid learning method bridging the fields of recurrent neural networks, unsupervised Hebbian learning, vector quantization, and supervised learning to implement a sophisticated image and feature segmentation architecture. This architecture is based on the competitive layer model (CLM), a dynamic feature binding model, which is applicable on a wide range of perceptual grouping and segmentation problems. A predefined target segmentation can be achieved as attractor states of this linear threshold recurrent network, if the lateral weights are chosen by Hebbian learning. The weight matrix is given by the correlation matrix of special pattern vectors with a structure dependent on the target labeling. Generalization is achieved by applying vector quantization on pair-wise feature relations, like proximity and similarity, defined by external knowledge. We show the successful application of the method to a number of artifical test examples and a medical image segmentation problem of fluorescence microscope cell images.

Decomposing a complex computational problem into sub-problems, which are computationally simpler to solve individually and which can be combined to produce a solution to the full problem, can efficiently lead to compact and general solutions. Modular neural networks represent one of the ways in which this divide-and-conquer strategy can be implemented. Here we present a co-evolutionary model which is used to design and optimize modular neural networks with task-specific modules. The model consists of two populations. The first population consists of a pool of modules and the second population synthesizes complete systems by drawing elements from the pool of modules. Modules represent a part of the solution, which co-operates with others in the module population to form a complete solution. With the help of two artificial supervised learning tasks created by mixing two sub-tasks we demonstrate that if a particular task decomposition is better in terms of performance on the overall task, it can be evolved using this co-evolutionary model.

A major problem in designing artificial neural networks is the proper choice of the network architecture. Especially for vision networks classifying three-dimensional (3-D) objects this problem is very challenging, as these networks are necessarily large and therefore the search space for defining the needed networks is of a very high dimensionality. This strongly increases the chances of obtaining only suboptimal structures from standard optimization algorithms. We tackle this problem in two ways. First, we use biologically inspired hierarchical vision models to narrow the space of possible architectures and to reduce the dimensionality of the search space. Second, we employ evolutionary optimization techniques to determine optimal features and nonlinearities of the visual hierarchy. Here, we especially focus on higher order complex features in higher hierarchical stages. We compare two different approaches to perform an evolutionary optimization of these features. In the first setting, we directly code the features into the genome. In the second setting, in analogy to an ontogenetical development process, we suggest the new method of an indirect coding of the features via an unsupervised learning process, which is embedded into the evolutionary optimization. In both cases the processing nonlinearities are encoded directly into the genome and are thus subject to optimization. The fitness of the individuals for the evolutionary selection process is computed by measuring the network classification performance on a benchmark image database. Here, we use a nearest-neighbor classification approach, based on the hierarchical feature output. We compare the found solutions with respect to their ability to generalize. We differentiate between a first- and a second-order generalization. The first-order generalization denotes how well the vision system, after evolutionary optimization of the features and nonlinearities using a database A, can classify previously unseen test views of objects from this database A. As second-order generalization, we denote the ability of the vision system to perform classification on a database B using the features and nonlinearities optimized on database A. We show that the direct feature coding approach leads to networks with a better first-order generalization, whereas the second-order generalization is on an equally high level for both direct and indirect coding. We also compare the second-order generalization results with other state-of-the-art recognition systems and show that both approaches lead to optimized recognition systems, which are highly competitive with recent recognition algorithms.

We introduce a decision making framework that uses evolutionary and learning methods. It is applied to competitive games to learn online the current opponent strategy and to adapt the system counter-strategy appropriately. We compared our system for the iterated prisoner's dilemma and rock-paper-scissors with three other methods against different typical game strategies as opponents. Results show that our system performs best in most cases and is able to adapt its strategy online to the current opponent. Moreover we could show that a good prediction of the opponent is no guaranty for a good payoff, since a good prediction is often the result of a poor opponent strategy which leads to a low payoff for both players.

Non-negative matrix factorization (NMF) is a method for the decomposition of multivariate data into strictly positive activations and basis vectors. Here, instead of using unstructured data vectors, we assume that something is known in advance about the type of transformations that either the input data or the basis vectors may undergo. This would be the case e.g. if we assume input vectors that are translationally shifted versions of each other, but it applies to any other transformations as well. The key idea is that we factorize the data into activations and basis vectors modulo the transformations. We show that this can be done by extending NMF in a natural way. The gained factorization thus provides a transformation-invariant and compact encoding that is optimal for the given transformation constraints.

There is an ongoing debate over the capabilities of hierarchical neural feedforward architectures for performing real-world invariant object recognition. Although a variety of hierarchical models exists, appropriate supervised and unsupervised learning methods are still an issue of intense research. We propose a feedforward model for recognition that shares components like weight sharing, pooling stages, and competitive nonlinearities with earlier approaches but focuses on new methods for learning optimal feature-detecting cells in intermediate stages of the hierarchical network. We show that principles of sparse coding, which were previously mostly applied to the initial feature detection stages, can also be employed to obtain optimized intermediate complex features. We suggest a new approach to optimize the learning of sparse features under the constraints of a weight-sharing or convolutional architecture that uses pooling operations to achieve gradual invariance in the feature hierarchy. The approach explicitly enforces symmetry constraints like translation invariance on the feature set. This leads to a dimension reduction in the search space of optimal features and allows determining more efficiently the basis representatives, which achieve a sparse decomposition of the input. We analyze the quality of the learned feature representation by investigating the recognition performance of the resulting hierarchical network on object and face databases. We show that a hierarchy with features learned on a single object data set can also be applied to face recognition without parameter changes and is competitive with other recent machine learning recognition approaches. To investigate the effect of the interplay between sparse coding and processing nonlinearities, we also consider alternative feedforward pooling nonlinearities such as presynaptic maximum selection and sum-of-squares integration. The comparison shows that a combination of strong competitive nonlinearities with sparse coding offers the best recognition performance in the difficult scenario of segmentation-free recognition in cluttered surround. We demonstrate that for both learning and recognition, a precise segmentation of the objects is not necessary.

We present a new approach to the supervised learning of lateral inter- actions for the competitive layer model (CLM) dynamic feature binding architecture. The method is based on consistency conditions, which were recently shown to characterize the attractor states of this linear threshold recurrent network. For a given set of training examples the learning prob- lem is formulated as a convex quadratic optimization problem in the lat- eral interaction weights. An efficient dimension reduction of the learning problem can be achieved by using a linear superposition of basis inter- actions. We show the successful application of the method to a medical image segmentation problem of fluorescence microscope cell images.

We establish two conditions that ensure the nondivergence of additive recurrent networks with unsaturating piecewise linear transfer functions, also called linear threshold or semilinear transfer functions. As Hahnloser, Sarpeshkar, Mahowald, Douglas, and Seung (2000) showed, networks of this type can be efficiently built in silicon and exhibit the coexistence of digital selection and analog amplification in a single circuit. To obtain this behavior, the network must be multistable and nondivergent, and our conditions allow determining the regimes where this can be achieved with maximal recurrent amplification. The first condition can be applied to nonsymmetric networks and has a simple interpretation of requiring that the strength of local inhibition match the sum over excitatory weights converging onto a neuron. The second condition is restricted to symmetric networks, but can also take into account the stabilizing effect of nonlocal inhibitory interactions. We demonstrate the application of the conditions on a simple example and the orientation-selectivity model of Ben-Yishai, Lev Bar-Or, and Sompolinsky (1995). We show that the conditions can be used to identify in their model regions of maximal orientation-selective amplification and symmetry breaking.

We present a recurrent neural network for feature binding and sensory segmentation: the competitive-layer model (CLM). The CLM uses topographically structured competitive and cooperative interactions in a layered network to partition a set of input features into salient groups. The dynamics is formulated within a standard additive recurrent network with linear threshold neurons. Contextual relations among features are coded by pairwise compatibilities, which define an energy function to be minimized by the neural dynamics. Due to the usage of dynamical winner-take-all circuits, the model gains more flexible response properties than spin models of segmentation by exploiting amplitude information in the grouping process. We prove analytic results on the convergence and stable attractors of the CLM, which generalize earlier results on winner-take-all networks, and incorporate deterministic annealing for robustness against local minima. The piecewise linear dynamics of the CLM allows a linear eigensubspace analysis, which we use to analyze the dynamics of binding in conjunction with annealing. For the example of contour detection, we show how the CLM can integrate figure-ground segmentation and grouping into a unified model.

We present the application of a recurrent neural network feature binding model to the segmentation of fluorescence micrographs, images showing fluorescent cells in tonsil tissue. Image primitives, referred to as features, consisting of position and local gradient information, build the input to the model. The competitive layer model is used to provide a binding of features to convex groups, corresponding to fluorescent cell bodies. Although the images contain noise, and the cells' shapes show considerable variation, the fluorescent cell contours are extracted with sufficient accuracy, according to a biomedical expert. The method achieves at the same time grouping and figure-ground segmentation, and does not require us to manually fix the number of groups.