Self-organizing Neural Architecture Research Articles

Unsupervised Machine Learning Based Scalable Fusion for Active Perception

The fusion of sensor data is an essential requirement for active perception in autonomous robots in order to accurately perceive an environment, make appropriate adjustments that improve the understanding of the environment and generate actionable insights. Despite a large body of literature in this domain, only a handful of research focuses on unsupervised machine learning, which is increasingly important in unlabeled, unknown environments. This paper proposes a scalable self-organizing neural architecture for environmental perception based on multimodal fusion using unsupervised machine learning. Inspired by the biological counterpart, the neural architecture consists of topographic maps across each modality and incorporates a novel scalable self-organization process that handles high volume, high-velocity sensor data. The co-occurrence relationships across modalities are captured in cross-modal connections and the reverse Hebbian projection is used for multimodal representation. This proposed unsupervised machine learning-based scalable fusion technique was implemented on Apache Hadoop and Apache Spark. It was validated using an extensive multimodal human activity sensor dataset to demonstrate efficient representation and multimodal fusion for active perception.

Compositional Learning of Human Activities With a Self-Organizing Neural Architecture.

An important step for assistive systems and robot companions operating in human environments is to learn the compositionality of human activities, i.e., recognize both activities and their comprising actions. Most existing approaches address action and activity recognition as separate tasks, i.e., actions need to be inferred before the activity labels, and are thus highly sensitive to the correct temporal segmentation of the activity sequences. In this paper, we present a novel learning approach that jointly learns human activities on two levels of semantic and temporal complexity: (1) transitive actions such as reaching and opening, e.g., a cereal box, and (2) high-level activities such as having breakfast. Our model consists of a hierarchy of GWR networks which process and learn inherent spatiotemporal dependencies of multiple visual cues extracted from the human body skeletal representation and the interaction with objects. The neural architecture learns and semantically segments input RGB-D sequences of high-level activities into their composing actions, without supervision. We investigate the performance of our architecture with a set of experiments on a publicly available benchmark dataset. The experimental results show that our approach outperforms the state of the art with respect to the classification of the high-level activities. Additionally, we introduce a novel top-down modulation mechanism to the architecture which uses the actions and activity labels as constraints during the learning phase. In our experiments, we show how this mechanism can be used to control the network's neural growth without decreasing the overall performance.

Lifelong learning of human actions with deep neural network self-organization

Lifelong learning is fundamental in autonomous robotics for the acquisition and fine-tuning of knowledge through experience. However, conventional deep neural models for action recognition from videos do not account for lifelong learning but rather learn a batch of training data with a predefined number of action classes and samples. Thus, there is the need to develop learning systems with the ability to incrementally process available perceptual cues and to adapt their responses over time. We propose a self-organizing neural architecture for incrementally learning to classify human actions from video sequences. The architecture comprises growing self-organizing networks equipped with recurrent neurons for processing time-varying patterns. We use a set of hierarchically arranged recurrent networks for the unsupervised learning of action representations with increasingly large spatiotemporal receptive fields. Lifelong learning is achieved in terms of prediction-driven neural dynamics in which the growth and the adaptation of the recurrent networks are driven by their capability to reconstruct temporally ordered input sequences. Experimental results on a classification task using two action benchmark datasets show that our model is competitive with state-of-the-art methods for batch learning also when a significant number of sample labels are missing or corrupted during training sessions. Additional experiments show the ability of our model to adapt to non-stationary input avoiding catastrophic interference.

Three-Dimensional Self-Organizing Neural Architectures: A Neural Stem Cells Reservoir and a System for Neurodevelopmental Studies

Complex microenvironmental stimuli influence neural cell properties. To study this, we developed a three-dimensional (3-D) neural culture system, composed of different populations including neurons, astrocytes, and neural stem cells (NSCs). In particular, these last-mentioned cells represent a source potentially exploitable to test drugs, to study neurodevelopment and cell-therapies for neuroregenerations. On seeding on matrigel in a medium supplemented with serum and mitogens, cells obtained from human fetal brain tissue formed 3-D self-organizing neural architectures. Immunocytochemical analysis demonstrated the presence of undifferentiated nestin+ and CD133+ cells, surrounded by β-tub-III+ and GFAP+ cells, suggesting the formation of niches containing potential human NSCs (hNSCs). The presence of hNSCs was confirmed by both neurosphere assay and RT-PCR, and their multipotentiality was demonstrated by both immunofluorescent staining and RT-PCR. Flow cytometry analysis revealed that neurosphere forming cells originating from at least two different subsets expressing, respectively, CD133 and CD146 markers were endowed with different proliferative and differentiation potential. Our data implicate that the complexity of environment within niches and aggregates of heterogeneous neural cell subsets may represent an innovative platform for neurobiological and neurodevelopmental investigations and a reservoir for a rapid expansion of hNSCs.

A self-organizing neural architecture integrating desire, intention and reinforcement learning

This paper presents a self-organizing neural architecture that integrates the features of belief, desire, and intention (BDI) systems with reinforcement learning. Based on fusion Adaptive Resonance Theory (fusion ART), the proposed architecture provides a unified treatment for both intentional and reactive cognitive functionalities. Operating with a sense-act-learn paradigm, the low level reactive module is a fusion ART network that learns action and value policies across the sensory, motor, and feedback channels. During performance, the actions executed by the reactive module are tracked by a high level intention module (also a fusion ART network) that learns to associate sequences of actions with context and goals. The intention module equips the architecture with deliberative planning capabilities, enabling it to purposefully maintain an agenda of actions to perform and to reduce the need of constantly sensing the environment. Through reinforcement learning, plans can also be evaluated and refined without the rigidity of user-defined plans. We examine two strategies for combining the intention and reactive modules for decision making in a real time environment. Our experiments based on a minefield navigation domain show that the integrated architecture is able to learn plans efficiently, achieve good plan utilization, and combine both intentional and reactive action execution to yield a robust performance.

Condition monitoring of the cutting process using a self-organizing spiking neural network map

This paper presents a new approach to sensor based condition monitoring using a self-organizing spiking neuron network map. Experimental evidence suggests that biological neural networks, which communicate through spikes, use the timing of these spikes to encode and compute information in a more efficient way. The paper introduces the basis of a simplified version of the Self-Organizing neural architecture based on Spiking Neurons. The fundamental steps for the development of this computational model are presented as well as some experimental evidence of its performance. It is shown that this computational architecture has a greater potential to unveil embedded information in tool wear monitoring data sets and that faster learning occurs if compared to traditional sigmoidal neural networks.

Self-organizing neural architecture and spiking neural model of the tadpole spinal cord

Event Abstract Back to Event Self-organizing neural architecture and spiking neural model of the tadpole spinal cord Roman Borisyuk1* 1 University of Plymouth, School of Computing and Mathematics, United Kingdom Traditional approach to a design of neural network model in computational neuroscience includes specification of the model's element, description of the architecture of connections between neurons, and formulation of the learning rule for modification of the connection strength. There is a broad range of biologically realistic models of a single neuron such as the conductance based Hodgkin-Huxley type models and their simplifications, e.g. Morris-Lecar model. As a rule in neuronal modelling, the architecture of neural connections is unknown. A typical approach is to use some standard connection scheme: all-to-all connections, local connections, random connections, feed-forward connections, etc. A more promising approach is to learn from a biological system how neurons self-organise their connectivity pattern to provide a proper functioning of the neural circuit.We use this approach in our collaborative project with Prof. Alan Roberts Lab, University of Bristol to study a neural network of spinal cord of young Xenopus tadpole (Li et al., 2007, Borisyuk et al., 2008). The spinal cord of tadpole is simple enough to allow detailed neurobiological experimentation with anatomy and neurophysiology of neurons. The spinal cord contains about 2000 neurons which can be arranged into six different cell types relevant to the control of swimming.Experimental data for this modelling have been provided by A. Roberts. In his Lab there is a unique collection of experimental data which includes both anatomical measurements on distribution of cell bodies, dendrites, axons in the spinal cord cylinder and multiple experimental data on electrophysiology of different neurons (Li et al., 2007).To define the connection architecture, we model a process of neural network development. A simple model of axon growth is a basis for the self-organisation process which results in the complete biologically realistic architecture of connections. Figure shows a fragment of reconstruction of the tadpole spinal cord. Each cell type is represented by a separate colour: axons, dendrites, initial point of axons, and synapses are shown. The dendrite is represented by a vertical bar and the synapse is shown by a circle. A colour of the circle relates to presynaptic neuron. More details are given in Borisyuk et al., 2008. The model includes both deterministic and stochastic components and allows to generate multiple copies of the spinal cord architecture with the same statistical characteristics of axon distribution as the measurements of real axons.We combine the generated anatomy with Morris-Lecar model of spiking neurons to find whether this neural network can produce a specific pattern of neural activity corresponding to the swimming of tadpole. Preliminary simulations show that the neural network with randomly distributed parameter values can swim. INCF-09-49

Self-Organizing Neural Architectures and Cooperative Learning in a Multiagent Environment

Temporal-Difference-Fusion Architecture for Learning, Cognition, and Navigation (TD-FALCON) is a generalization of adaptive resonance theory (a class of self-organizing neural networks) that incorporates TD methods for real-time reinforcement learning. In this paper, we investigate how a team of TD-FALCON networks may cooperate to learn and function in a dynamic multiagent environment based on minefield navigation and a predator/prey pursuit tasks. Experiments on the navigation task demonstrate that TD-FALCON agent teams are able to adapt and function well in a multiagent environment without an explicit mechanism of collaboration. In comparison, traditional Q-learning agents using gradient-descent-based feedforward neural networks, trained with the standard backpropagation and the resilient-propagation (RPROP) algorithms, produce a significantly poorer level of performance. For the predator/prey pursuit task, we experiment with various cooperative strategies and find that a combination of a high-level compressed state representation and a hybrid reward function produces the best results. Using the same cooperative strategy, the TD-FALCON team also outperforms the RPROP-based reinforcement learners in terms of both task completion rate and learning efficiency.

A VLSI architecture for video-object segmentation

Object-based coding and description in real time are increasingly important for many image and video applications. We propose a very large scale integration (VLSI) architecture based on a novel segmentation algorithm for extracting objects in video. The segmentation architecture of a frame mainly consists of two functional phases. In the first phase, pixel-based static and dynamic features in video are extracted, filtered, normalized. These multiple features are labeled using a self-organizing feature map neural networks architecture to generate initial segmentation labels. An edge fusion module in the second phase combines the initial segmentation labels and a linked edge map of a frame to generate more accurate segmentation where region boundaries are closer to the actual object boundaries. Computational and hardware complexities of the proposed architecture are estimated in terms of the number of clocks, arithmetic components, gates, and memory requirements. The performance of the proposed VLSI architecture demonstrates the possibility of performing object-oriented video segmentation in real time.

Deformation invariant visual object recognition: Experiments with a self-organising neural architecture

A self-organising neural network architecture for grey-scale visual object rcognition is presented. The network is composed of three processing layers with an architecture designed to give deformation tolerance. The processing layers involve feature extraction, sub-pattern detection and classification. Training is generally performed on-line in an unsupervised manner, classes being created when objects are presented that cannot be classified. The results given show the effect of the two discrimination parameters when the network is applied to two very different sets of images, namely hand written numerals and hand gestures images. The sensitivity of the network to the parameters that govern the size of detectable patterns and the areas over which they are detected is also tested. The robustness of the network to the order of image presentation is also demonstrated. The results show that parameter choice is not critical and heuristically chosen parameters provide near optimum performance.

Fast-learning VIEWNET architectures for recognizing three-dimensional objects from multiple two-dimensional views

The recognition of three-dimensional (3-D) objects from sequences of their two-dimensional (2-D) views is modeled by a family of self-organizing neural architectures, called VIEWNET, that use View Information Encoded With NETworks. VIEWNET incorporates a preprocessor that generates a compressed but 2-D invariant representation of an image, a supervised incremental learning system that classifies the preprocessed representations into 2-D view categories whose outputs are combined into 3-D invariant object categories, and a working memory that makes a 3-D object prediction by accumulating evidence from 3-D object category nodes as multiple 2-D views are experienced. The simplest VIEWNET achieves high recognition scores without the need to explicitly code the temporal order of 2-D views in working memory. Working memories are also discussed that save memory resources by implicitly coding temporal order in terms of the relative activity of 2-D view category nodes, rather than as explicit 2-D view transitions. Variants of the VIEWNET architecture may be used for scene understanding by using a preprocessor and classier that can determine both what objects are in a scene and where they are located. The present VIEWNET preprocessor includes the CORT-X 2 filter, which discounts the illuminant, regularizes and completes figural boundaries, and suppresses image noise. This boundary segmentation is rendered invariant under 2-D translation, rotation, and dilation by use of a log polar transform. The invariant spectra undergo Gaussian coarse coding to further reduce noise and 3-D foreshortening effects, and to increase generalization. These compressed codes are input into the classifier, a supervised learning system based on the fuzzy ARTMAP algorithm. Fuzzy ARTMAP learns 2-D view categories that are invariant under 2-D image translation, rotation, and dilation as well as 3-D image transformations that do not cause a predictive error. Evidence from sequences of 2-D view categories converges at 3-D object nodes that generate a response invariant under changes of 2-D view. These 3-D object nodes input to a working memory that accumulates evidence over time to improve object recognition. In the simplest working memory, each occurrence (nonoccurrence) of a 2-D view category increases (decreases) the corresponding node's activity in working memory. The maximally active node is used to predict the 3-D object. Recognition is studied with noisy and clean images using slow and fast learning. Slow learning at thefuzzy ARTMAP mapfield is adapted to learn the conditional probability of the 3-D object given the selected 2-D view category. VIEWNET is demonstrated on an MIT Lincoln Laboratory database of 128 × 128 2-D views of aircraft with and without additive noise. A recognition rate of up to 90% is achieved with one 2-D view and of up to 98.5% correct with three 2-D views. The properties of 2-D view and 3-D object category nodes are compared with those of cells in monkey inferotemporal cortex.