When working in a group, it is essential to understand each other’s viewpoints to increase group cohesion and meeting productivity. This can be challenging in teams: participants might be left misunderstood and the discussion could be going around in circles. To tackle this problem, previous research on group interactions has addressed topics such as dominance detection, group engagement, and group creativity. Conversational memory, however, remains a widely unexplored area in the field of multimodal analysis of group interaction. The ability to track what each participant or a group as a whole find memorable from each meeting would allow a system or agent to continuously optimise its strategy to help a team meet its goals. In the present paper, we therefore investigate what participants take away from each meeting and how it is reflected in group dynamics.As a first step toward such a system, we recorded a multimodal longitudinal meeting corpus (MEMO), which comprises a first-party annotation of what participants remember from a discussion and why they remember it. We investigated whether participants of group interactions encode what they remember non-verbally and whether we can use such non-verbal multimodal features to predict what groups are likely to remember automatically. We devise a coding scheme to cluster participants’ memorisation reasons into higher-level constructs. We find that low-level multimodal cues, such as gaze and speaker activity, can predict conversational memorability. We also find that non-verbal signals can indicate when a memorable moment starts and ends. We could predict four levels of conversational memorability with an average accuracy of 44 %. We also showed that reasons related to participants’ personal feelings and experiences are the most frequently mentioned grounds for remembering meeting segments.

Pixelwise annotation of image sequences can be very tedious for humans. Interactive video object segmentation aims to utilize automatic methods to speed up the process and reduce the workload of the annotators. Most contemporary approaches rely on deep convolutional networks to collect and process information from human annotations throughout the video. However, such networks contain millions of parameters and need huge amounts of labeled training data to avoid overfitting. Beyond that, label propagation is usually executed as a series of frame-by-frame inference steps, which is difficult to be parallelized and is thus time consuming. In this paper we present a graph neural network based approach for tackling the problem of interactive video object segmentation. Our network operates on superpixel-graphs which allow us to reduce the dimensionality of the problem by several magnitudes. We show, that our network possessing only a few thousand parameters is able to achieve state-of-the-art performance, while inference remains fast and can be trained quickly with very little data.

Deep Neural Networks (DNNs) achieve the state-of-the-art results on a wide range of image processing tasks, however, the majority of such solutions are problem-specific, like most AI algorithms. The One Network to Solve Them All (OneNet) procedure has been suggested to resolve this issue by exploiting a DNN as the proximal operator in Alternating Direction Method of Multipliers (ADMM) solvers for various imaging problems. In this work, we make two contributions, both facilitating end-to-end learning using backpropagation. First, we generalize OneNet to videos by augmenting its convolutional prior network with bidirectional recurrent connections; second, we extend the fixed fully connected linear ADMM data step with another trainable bidirectional convolutional recurrent network. In our computational experiments on the Rotated MNIST, Scanned CIFAR-10 and UCF-101 data sets, the proposed modifications improve performance by a large margin compared to end-to-end convolutional OneNet and 3D Wavelet sparsity on several video processing problems: pixelwise inpainting-denoising, blockwise inpainting, scattered inpainting, super resolution, compressive sensing, deblurring, frame interpolation, frame prediction and colorization. Our two contributions are complementary, and using them together yields the best results.

Manual annotation of video segmentation datasets requires an immense amount of human effort, thus, reduction of human annotation costs is an active topic of research. While many papers deal with the propagation of masks through frames of a video, only a few results attempt to optimize annotation task selection. In this paper we present a deep learning based solution to the latter problem and train it using Reinforcement Learning. Our approach utilizes a modified version of the Dueling Deep Q-Network sharing weight parameters across the temporal axis of the video. This technique enables the trained agent to select annotation tasks from the whole video. We evaluate our annotation task selection method by means of a hierarchical supervoxel segmentation based mask propagation algorithm.

Medial axis representation (a.k.a. shape skeleton) seems to be present in visual processing, but its relevance has remained unclear. Here, we show the potentials of the medial axis transformation in the temporal propagation of superpixels. We combine (i) state-of-the-art deep neural network `sensors' for optical flow and for depth estimation and (ii) a superpixel algorithm with (iii) the medial axis transformation to obtain frame-to-frame propagation of visual objects. We study the precision of this deep learning facilitated superpixel temporal propagation. We discuss the advantages of the method compared to the temporal propagation of the superpixels themselves.

We estimated the contribution of different factors in segmentation tasks by means of deep neural networks. Results indicated that texture and optical flow have similar power, but they seem not to add up. In turn, we decided to study the `Common Fate Principle' of the 100 years gestaltism suggesting that elements that move together belong together. We developed a simple, fast, and efficient episodic segmentation method that - to some extent - resembles the `how system' of the visual processing: we dropped every piece of information except motion, and started from pure optical flow estimations on 2D videos. For the sake of segmentation, we used a parallel and fast hierarchical supervoxel algorithm. We studied (i) grid topology in space and time, (ii) 2D grid in space and topology dictated by the optical flow in time, and (iii) added deep network based depth estimation from 2D images. We measure performances on episodic foreground-background segmentation task of the Davis benchmark videos. Results are competitive to state-of-the-art segmentation techniques.

In this paper we propose Group k-Sparse Temporal Convolutional Neural Networks for unsupervised pretraining using video data. Our work is the first to consider the recurrent extension of structured sparsity, thus enhancing representational power and explainability. We show that our architecture is able to outperform several state-of-the-art baselines on Rotated MNIST, Scanned CIFAR-10, COIL-100 and NEC Animal pretraining benchmarks for video classification using limited labeled data.

The common approach to 3D human pose estimation is predicting the body joint coordinates relative to the hip. This works well for a single person but is insufficient in the case of multiple interacting people. Methods predicting absolute coordinates first estimate a root-relative pose then calculate the translation via a secondary optimization task. We propose a neural network that predicts joints in a camera centered coordinate system instead of a root-relative one. Unlike previous methods, our network works in a single step without any post-processing. Our network beats previous methods on the MuPoTS-3D dataset and achieves state-of-the-art results.

In this paper, we describe how we combine active and passive user input modes in clinical environments for knowledge discovery and knowledge acquisition towards decision support in clinical environments. Active input modes include digital pens, smartphones, and automatic handwriting recognition for a direct digitalisation of patient data. Passive input modes include sensors of the clinical environment and or mobile smartphones. This combination for knowledge acquisition and decision support (while using machine learning techniques) has not yet been explored in clinical environments and is of specific interest because it combines previously unconnected information sources for individualised treatments. The innovative aspect is a holistic view on individual patients based on ontologies, terminologies, and textual patient records whereby individual active and passive real-time patient data can be taken into account for improving clinical decision support.

Power Factor Correction (PFC) converters belong to the variable structure piecewise linear systems due to the switching action. Their complex behaviour is intensively studied. Present paper represents a stability analysis method using the socalled auxiliary state vector to determine the Jacobian matrix of the Poincare Map Function (PMF) to calculate the critical angle where oscillations developes in the inductor current. The effect caused by the Zero-Order Hold, the digital computation delays and the nonideal circuit elements are taken into consideration. Furthermore by adding a stabilizing signal the stable range can be extended. The parameters of the stabilizing signal is also determined by the auxiliary state vector. The theoretic results are verified by simulation and the experiment results.

Sensory representations are not only sparse, but often overcomplete: coding units significantly outnumber the input units. For models of neural coding this overcompleteness poses a computational challenge for shaping the signal processing channels as well as for using the large and sparse representations in an efficient way. We argue that higher level overcompleteness becomes computationally tractable by imposing sparsity on synaptic activity and we also show that such structural sparsity can be facilitated by statistics based decomposition of the stimuli into typical and atypical parts prior to sparse coding. Typical parts represent large-scale correlations, thus they can be significantly compressed. Atypical parts, on the other hand, represent local features and are the subjects of actual sparse coding. When applied on natural images, our decomposition based sparse coding model can efficiently form overcomplete codes and both center-surround and oriented filters are obtained similar to those observed in the retina and the primary visual cortex, respectively. Therefore we hypothesize that the proposed computational architecture can be seen as a coherent functional model of the first stages of sensory coding in early vision.

We develop a dictionary learning method which is (i) online, (ii) enables overlapping group structures with (iii) non-convex sparsity-inducing regularization and (iv) handles the partially observable case. Structured sparsity and the related group norms have recently gained widespread attention in group-sparsity regularized problems in the case when the dictionary is assumed to be known and fixed. However, when the dictionary also needs to be learned, the problem is much more difficult. Only a few methods have been proposed to solve this problem, and they can handle two of these four desirable properties at most. To the best of our knowledge, our proposed method is the first one that possesses all of these properties. We investigate several interesting special cases of our framework, such as the online, structured, sparse non-negative matrix factorization, and demonstrate the efficiency of our algorithm with several numerical experiments.

In this paper we propose an algorithm for polynomial-time reinforcement learning in factored Markov decision processes (FMDPs). The factored optimistic initial model (FOIM) algorithm, maintains an empirical model of the FMDP in a conventional way, and always follows a greedy policy with respect to its model. The only trick of the algorithm is that the model is initialized optimistically. We prove that with suitable initialization (i) FOIM converges to the fixed point of approximate value iteration (AVI); (ii) the number of steps when the agent makes non-near-optimal decisions (with respect to the solution of AVI) is polynomial in all relevant quantities; (iii) the per-step costs of the algorithm are also polynomial. To our best knowledge, FOIM is the first algorithm with these properties.

The hippocampal formation is believed to play a central role in forming long lasting representation of events. However, in contrast to the continuous nature of sensory signal flow, events are spatially and temporally bounded processes. In this paper we are interested in the kind of representation that allows for detecting and/or predicting events. Based on new results on the identification problem of linear hidden processes, we propose a general signal encoding model that can represent causal relationships used to define events. We translate the model into a connectionist structure in which parameter learning follows biologically plausible rules. We also speculate on the resemblance of the resulting structure to the connection system of the hippocampal formation. When our signal encoding model is applied on spatially anchored inputs, its different parts feature spatially localized and periodic neural activity similar to those found in the hippocampus and in the entorhinal cortex, respectively. These emergent forms of spatial activity differentiates our model from other computational models of (spatial) memory as the model has not been explicitly designed to deal with spatial information. We speculate that our model may describe the core function of the hippocampal region in forming episodic memory and supporting spatial navigation.

In this paper we address the controlled complete AutoRegressive Moving Average Independent Process Analysis (ARMAX-IPA; X-exogenous input or control) problem, which is a generalization of the Blind SubSpace Deconvolution (BSSD) task. Compared to our previous work that dealt with the undercomplete situation, (i) here we extend the theory to complete systems, (ii) allow an autoregressive part to be present, (iii) and include exogenous control. We investigate the case when the observed signal is a linear mixture of independent multidimensional ARMA processes that can be controlled. Our objective is to estimate the ARMA processes, their driving noises as well as the mixing. We aim efficient estimation by choosing suitable control values. For the optimal choice of the control we adapt the D-optimality principle, also known as the ‘InfoMax method’. We solve the problem by reducing it to a fully observable D-optimal ARX task and Independent Subspace Analysis (ISA) that we can solve. Numerical examples illustrate the efficiency of the proposed method.

The exploration-exploitation dilemma has been an intriguing and unsolved problem within the framework of reinforcement learning. "Optimism in the face of uncertainty" and model building play central roles in advanced exploration methods. Here, we integrate several concepts and obtain a fast and simple algorithm. We show that the proposed algorithm finds a near-optimal policy in polynomial time, and give experimental evidence that it is robust and efficient compared to its ascendants.

In this article we propose a method that can deal with certain combinatorial reinforcement learning tasks. We demonstrate the approach in the popular Ms. Pac-Man game. We define a set of high-level observation and action modules, from which rule-based policies are constructed automatically. In these policies, actions are temporally extended, and may work concurrently. The policy of the agent is encoded by a compact decision list. The components of the list are selected from a large pool of rules, which can be either hand-crafted or generated automatically. A suitable selection of rules is learnt by the cross-entropy method, a recent global optimization algorithm that fits our framework smoothly. Cross-entropy-optimized policies perform better than our hand-crafted policy, and reach the score of average human players. We argue that learning is successful mainly because (i) policies may apply concurrent actions and thus the policy space is sufficiently rich, (ii) the search is biased towards low-complexity policies and therefore, solutions with a compact description can be found quickly if they exist.

The cross-entropy method is an efficient and general optimization algorithm. However, its applicability in reinforcement learning (RL) seems to be limited because it often converges to suboptimal policies. We apply noise for preventing early convergence of the cross-entropy method, using Tetris, a computer game, for demonstration. The resulting policy outperforms previous RL algorithms by almost two orders of magnitude.

A novel algorithm for performing Independent Subspace Analysis, the estimation of hidden independent subspaces is introduced. This task is a generalization of Independent Component Analysis. The algorithm works by estimating the multi-dimensional differential entropy. The estimation utilizes minimal geodesic spanning trees matched to the sample points. Numerical studies include (i) illustrative examples, (ii) a generalization of the cocktail-party problem to songs played by bands, and (iii) an example on mixed independent subspaces, where subspaces have dependent sources, which are pairwise independent.

Gathering of novel information from the WWW constitutes a real challenge for artificial intelligence (AI) methods. Large search engines do not offer a satisfactory solution, their indexing cycle is long and they may offer a huge amount of documents. An AI-based link-highlighting procedure designed to assist surfing is studied here. It makes use of (i) 'experts', i.e. pretrained classifiers, forming the long-term memory of the system, (ii) relative values of experts and value estimation of documents based on recent choices of the users. Value estimation adapts fast and forms the short-term memory of the system. All experiments show that surfing based filtering can efficiently highlight 10-20% of the documents in about 5 steps, or less.

We studied user behaviors when the cursor is directed by a head in a simple control task. We used an intelligent writing tool called Dasher. Hidden Markov models (HMMs) were applied to separate behavioral patterns. We found that similar interpretations can be given to the hidden states upon learning. It is argued that the recognition of such general application specific behavioral patterns should be of help for adaptive human-computer interfaces.

We populated a huge scale-free portion of Internet environment with news foragers. They evolved by a simple internal selective algorithm: selection concerned the memory components, being finite in size and containing the list of most promising supplies. Foragers received reward for locating not yet found news and used value estimation. Foragers were allowed to multiply if they passed a given productivity threshold. A particular property of this community is that there is no direct interaction (here, communication) amongst foragers. It is found that, still, fast compartmentalization, i.e., fast division of work can be achieved.

In this paper we study the emergent structure of networks in which spike-timing dependent synaptic plasticity is induced only by external random noise. We show that such noise driven Hebbian networks are able to develop a broad range of network structures, including scale-free small-world networks. The development of such network structures may provide an explanation of the role of noise and its interplay with Hebbian plasticity. We also argue that this model can be seen as a unification of the famous Watts-Strogatz and preferential attachment models of small-world and scale-free nets. Our results may support Edelman's idea on that the development of central nervous system may have evolutionary components.

We can memorize long sequences like melodies or poems and it is intriguing to develop efficient connectionist representations for this problem. Recurrent neural networks have been proved to offer a reasonable approach here. We start from a few axiomatic assumptions and provide a simple mathematical framework that encapsulates the problem. A gradient-descent based algorithm is derived in this framework. Demonstrations on a benchmark problem show the applicability of our approach.

For interacting agents in time-critical applications, learning whether a subtask can be scheduled reliably is an important issue. The identification of sub-problems of this nature may promote e.g., planning, scheduling and segmenting in Markov decision processes. We define a subtask to be schedulable if its execution time has a small variance. We present an algorithm for finding such subtasks.

There is a growing interest in using Kalman filter models in brain modeling. The question arises whether Kalman filter models can be used on-line not only for estimation but for control. The usual method of optimal control of Kalman filter makes use of off-line backward recursion, which is not satisfactory for this purpose. Here, it is shown that a slight modification of the linear-quadratic-gaussian Kalman filter model allows the on-line estimation of optimal control by using reinforcement learning and overcomes this difficulty. Moreover, the emerging learning rule for value estimation exhibits a Hebbian form, which is weighted by the error of the value estimation.

In optimizations the dimension of the problem may severely, sometimes exponentially increase optimization time. Parametric function approximatiors (FAPPs) have been suggested to overcome this problem. Here, a novel FAPP, cost component analysis (CCA) is described. In CCA, the search space is resampled according to the Boltzmann distribution generated by the energy landscape. That is, CCA converts the optimization problem to density estimation. Structure of the induced density is searched by independent component analysis (ICA). The advantage of CCA is that each independent ICA component can be optimized separately. In turn, (i) CCA intends to partition the original problem into subproblems and (ii) separating (partitioning) the original optimization problem into subproblems may serve interpretation. Most importantly, (iii) CCA may give rise to high gains in optimization time. Numerical simulations illustrate the working of the algorithm.

We present a prototype of a recently proposed two stage model of the entorhinal-hippocampal loop. Our aim is to form a general computational model of the sensory neocortex. The model--grounded on pure information theoretic principles--accounts for the most characteristic features of long-term memory (LTM), performs bottom-up novelty detection, and supports noise filtering. Noise filtering can also serve to correct the temporal ordering of information processing. Surprisingly, as we examine the temporal characteristics of the model, the emergent dynamics can be interpreted as perceptual priming, a fundamental type of implicit memory. In the model's framework, computational results support the hypothesis of a strong correlation between perceptual priming and repetition suppression and this correlation is a direct consequence of the temporal ordering in forming the LTM. We also argue that our prototype offers a relatively simple and coherent explanation of priming and its relation to a general model of information processing by the brain.

A functional model of the basal ganglia-thalamocortical (BTC) loops is described. In our modeling effort, we try to minimize the complexity of our starting hypotheses. For that reason, we call this type of modeling Ockham's razor modeling. We have the additional constraint that the starting assumptions should not contradict experimental findings about the brain. First assumption: The brain lacks direct representation of paths but represents directions (called speed fields in control theory). Then control should be concerned with speed-field tracking (SFT). Second assumption: Control signals are delivered upon differencing in competing parallel channels of the BTC loops. This is modeled by extending SFT with differencing that gives rise to the robust Static and Dynamic State (SDS) feedback-controlling scheme. Third assumption: Control signals are expressed in terms of a gelatinous medium surrounding the limbs. This is modeled by expressing parameters of motion in parameters of the external space. We show that corollaries of the model fit properties of the BTC loops. The SDS provides proper identification of motion related neuronal groups of the putamen. Local minima arise during the controlling process that works in external space. The model explains the presence of parallel channels as the means to avoiding such local minima. Stability conditions of the SDS predict that the initial phase of learning is mostly concerned with selection of sign for the inverse dynamics. The model provides a scalable controller. State description in external space instead of configurational space reduces the dimensionality problem. Falsifying experiment is suggested. Computer experiments demonstrate the feasibility of the approach. We argue that the resulting scheme has a straightforward connectionist representation exhibiting population coding and Hebbian learning properties.

This paper presents a generative data reconstruction neural network model equipped with plastic lateral connections. The model is capable of capturing basic phenomena related to category formation. It explains category formation as an effect of cumulative memory traces at the level of lateral connectivity. The formed memory traces change network activity that is the basis of categorization according to the model. This change depends on the structure of the lateral connectivity and on the stimuli used in demonstrations. We argue that the model resolves the seemingly contradictory demonstrational results carried out with Alzheimer disease patients on category formation. We consider different stimulus sets and degraded lateral connectivity and show that the categorization probability can change from monotone to non-monotone functions depending on the sets.

We show that several properties of the highly elaborate structure of the EC-HC loop can be explained using the single principle that to recall past and to foresee future events a predictive structure is necessary. Networks that develop independent components (ICs) in an efficient manner can be built from two stages. We identify, these stages with the CA3 and CAI layers of the hippocampus (HC). The forming of ICs requires nonlinear operation, whereas IC outputs arise under linear operation and thus two-phase operation follows. Concurrent occurrences of past and present events are required by Hebbian learning and can be achieved by delaying structures. The loop structure requires a third layer that we identify with the entorhinal cortex (EC). Proper encoding into the EC is possible during linear operation in a supervised manner. The DRN can be seen as an error compensating control architecture. Thus the novel information processed by the DRN may be temporally convolved. We assume that blind source deconvolution is executed by the dentate gyrus and show that the dentate gyrus can satisfy the requirements.

It is argued that a novel control architecture, the Static and Dynamic State (SDS) feedback scheme, which utilizes speed-field tracking, exhibits global stability, and allows on-line tuning by any adaptation mechanism without canceling stability if certain structural conditions are met, can be viewed as a model of basal ganglia-thalamocortical loops since (1) the SDS scheme predicts the neuronal groups that fit neuronal classification in the supplementary motor area, the motor cortex and the putamen, (2) the structural stability conditions require parallel channels, a feature that these loops provide, and (3) the SDS scheme predicts two major disorders that can be identified as Parkinson's and Huntington's diseases. Simulations suggests that the basal ganglia work outside the realm of the stability condition allowed by the robustness of the scheme and required for increased computation speeds.

It is shown that local, extended objects of a metrical topological space shape the receptive fields of competitive neurons to local filters. Self-organized topology learning is then solved with the help of Hebbian learning together with extended objects that provide unique information about neighborhood relations. A topographical map is deduced and is used to speed up further adaptation in a changing environment with the help of Kohonen-type learning that teaches the neighbors of winning neurons as well.