This work builds on a cognitive theory called dynamic logic and considers the relationship between language and cognition. We explore the idea of dual models that combine linguistic and sensor features. We demonstrate that simultaneous learning of textual and image data results in formation of meaningful concepts and subsequent improvement in concept recognition.
In a decision study called the Asian Disease Problem, Tversky and Kahneman [1] found that framing risky health choices in terms of gains or losses of lives leads to radically different choices: risk seeking for losses and risk avoidance for gains. The difference between the two choices is called the framing effect. The authors explained framing effects via psychophysics of the numbers of lives saved or lost. Yet Reyna and Brainerd [2] showed that the strength of the framing effect depended not on the numbers but on whether one of options explicitly contained the possibility of no lives lost or saved. They fit their explanation into fuzzy trace theory whereby decisions are based not on details of the options given but on the gist (underlying meaning) of the options. We discuss how a brain-based neural network model of other decision data [3] that combines fuzzy trace theory with adaptive resonance theory can be extended to these framing data. Simulations are in progress.
We use the term "neurocognitive architecture" here to refer to any artificially intelligent agent where cognitive functions are implemented using brain-inspired neurocomputational methods. Creating and studying neurocognitive architectures is a very active and increasing focus of research efforts. We have recently been exploring the use of neural activity limit cycles as representations of perceived external information in self-organizing maps (SOMs). Specifically, we have been examining limit cycle representations in terms of their compatibility with self-organizing map formation and as working memory encodings for cognitively-relevant stimuli (e.g., for images of objects and their corresponding names expressed as phoneme sequences [1]). Here we evaluate the use of limit cycle representations in a new context of relevance to any cognitive agent: representing a spatial location. We find that, following repeated exposure to external 2D coordinate input values, robust limit cycles occur in a network's map region, the limit cycles representing nearby locations in external space are close to one another in activity state space, and the limit cycles representing widely separated external locations are very different from one another. Further, and in spite of the continually varying activity patterns in the network (instead of the fixed activity patterns used in most SOM work), map formation based on the learned limit cycles still occurs. We believe that these results, along with those in our earlier work, make limit cycle representations potentially useful for encoding information in the working memory of neurocognitive architectures.
This paper describes a novel distributed quasi-Newton-based robust training using the normalized risk-averting error (NRAE) with the gradual deconvexification (GDC) strategy. The main purpose of the computation is accomplished by optimizing the NRAE criterion parallely across different computing units, thereby two big advantages such as faster computation and global convergence can be obtained. The key idea is to replace the log partition function of the NRAE with a parallelizable upper-bound based on the concavity of the log-function. As a result, it is confirmed that the method is robust, and provides high quality training solutions regardless of initial values. Furthermore, the CPU time is drastically improved by the proposed distribution method without losing the quality of solutions.
In this work we describe how an existing neural model for learning Cell Assemblies (CAs) across multiple neuroanatomical brain areas has been integrated with a humanoid robot simulation to explore the learning of associations of visual and motor modalities. The results show that robust CAs are learned to enable pattern completion to select a correct motor response when only visual input is presented. We also show, with some parameter tuning and the pre-processing of more realistic patterns taken from images of real objects and robot poses the network can act as a controller for the robot in visuo-motor association tasks. This provides the basis for further neurorobotic experiments on grounded language learning.
Among different languages' sentence completion and parsing, Chinese is of great difficulty. Chinese words are not naturally separated by delimiters, which imposes extra challenge. Cogent confabulation based sentence completion has been proposed for English. It fills in missing words in an English sentence while maintains the semantic and syntactic consistency. In this work, we improve the cogent confabulation model and apply it to sentence completion in Chinese. Incorporating trained knowledge in parts-of-speech tagging and Chinese word compound segmentation, the model does not only fill missing words in a sentence but also performs linguistic analysis of the sentence with a high accuracy. We further investigate the optimization of the model and trade-offs between accuracy and training/recall complexity. Experimental results show that the optimized model improves recall accuracy by 9% and reduces training and recall time by 18.6% and 53.7% respectively.
Hierarchical Temporal Memory (HTM) is a model with hierarchically connected modules doing spatial and temporal pattern recognition, as described by Jeff Hawkins in his book entitled On Intelligence. Cortical Learning Algorithms (CLAs) comprise the second implementation of HTM. CLAs are an attempt by Numenta Inc. to create a computational model of perceptual analysis and learning inspired by the neocortex in the brain. In its current state only an implementation of one isolated region has been completed. The goal of this paper is to test if adding a second higher level region implementing CLAs to a system with just one region of CLAs, helps in improving the prediction accuracy of the system. The LIDA model (Learning Intelligent Distribution Agent - LIDA is a cognitive architecture) can use such a hierarchical implementation of CLAs for its Perceptual Associative Memory.
In this paper, we report the results of a study on the post-intervention effects of applying anodal transcranial Direct Current Stimulation (tDCS) on the intensity of motor Event Related Desynchronization. Ten subjects were given 15 minutes of sham and 1.5 mA tDCS on two separate occasions in randomized order in a double blind setting. Post-intervention EEG was then recorded while subjects were asked to perform imagined motor imagery. Results show that the intensity of 8-13Hz Mu rhythms exhibited significant difference between the sham and tDCS groups, with an average of 24.13 μV2 for sham and 32.57 μV2 for tDCS with a measured t-test p value of 0.03.
This paper presents a simulation model for observing the dynamics of a research community in engineering. The objective is to study how parameters, like group expertise and resources, influence the effectiveness of the group and community as a whole. The model implements a game-theoretic approach in which every group maximizes the difference between its rewards and costs (e.g., time and resources). Experiments studied the total reward and the number of problems solved over time by a community made of twenty groups for different conditions, i.e. allocated resources and group characteristics.
Decision-making is an interdisciplinary problem that has been the focus of many studies, particularly by interactive pairs of visual cognition systems. In a series of experiments, Bahrami et al. (2010) showed that dyadic interaction is beneficial only if participants communicate with each other about their confidence in making a judgment. Aside from data combination using both simple and weighted average, Hsu et al. (2006) first described the use of combinatorial fusion to combine multiple scoring systems (MSS). In this experiment, sixteen trials were conducted using pairs of individuals as visual cognition systems. Participants observed a target being thrown in a grassy field which could not be seen once it had landed, allowing them to then independently perceive the position of the target and determine their confidence level. The results of these trials were analyzed for performance of score and rank combinations relative to both the original cognition systems of the individuals and to simple and weighted averages of their systems. We demonstrated, using combinatorial fusion, that the combination of two visual perception systems is better than each of the individual systems only if they perform relatively well and they are diverse.
With a focus on presenting information at the right time, the ubicomp community can benefit greatly from learning the most salient human measures of cognitive load. Cognitive load can be used as a metric to determine when or whether to interrupt a user. In this paper, we collected data from multiple sensors and compared their ability to assess cognitive load. Our focus is on visual perception and cognitive speed-focused tasks that leverage cognitive abilities common in ubicomp applications. We found that across all participants, the electrocardiogram median absolute deviation and median heat flux measurements were the most accurate at distinguishing between low and high levels of cognitive load, providing a classification accuracy of over 80% when used together. Our contribution is a real-time, objective, and generalizable method for assessing cognitive load in cognitive tasks commonly found in ubicomp systems and situations of divided attention.
An experimental design is programmed using the presentation tool to investigate the global response inhibition process by quantifying the parameters such as inhibition efficiency, stop-signal delay (SSD) and stop-signal reaction time (SSRT) in the stop-signal paradigm. The aim of this study is to explore the response inhibition mechanisms in the left-hand and right-hand responses by using ERP and ERSP results obtained from the EEG data of different subjects. The inhibition efficiency of the right-hand response and left-hand response appears to be independent of each other as there is no significant difference between them. From these results, the inhibition mechanisms corresponding to these two regions of the brain may be viewed as statistically independent processes. Further, we inferred that the response inhibition mechanisms for both left-hand and right-hand responses have approximately the same spectral power observation analysis and we conclude that these processes are statistically independent of each other.
More convincing evidence has proven the existence of a bidirectional relationship between neurons and astrocytes. Assume now that astrocytes, a new type of glial cells previously considered as passive cells of support, constitute a system of nonsynaptic transmission plays a major role in modulating the activity of neurons. In this context, we proposed to model the effect of these cells to develop a new type of artificial neural network operating on new mechanisms to improve the information processing and reduce learning time, very expensive in traditional networks. The obtained results indicate that the implementation of bio-inspired functions such as of astrocytes, improve very considerably learning speed. The developed model achieves learning up to twelve times faster than traditional artificial neural networks.
We propose in this paper a novel approach of adaptive filtering of EEG signals. The filter adapts to the intrinsic characteristics of each person. The goal of the proposed method is to enhance the accuracy of the home devices system controlled by the thoughts related to two motor imagery actions. μ-rhythm and β-rhythm are the specific returned bands that contain the information. The main idea of the proposed method is to preserve the frequency bands of interest with a different value of the SNR on the stop-band. Our experimental results show the benefits of a suitable tuning of the filter on the accuracy of the classifier on the output of the EEG system. The proposed approach outperforms significantly performances reported in the literature and the effectively enhancement of the classification accuracy can reach up to 40% based only on filtering tuning.
The premiss that a group of cooperating agents - a collective brain - can solve a problem more efficiently than the same group of agents working independently is widespread, despite the little quantitative groundwork to support it. Here we use extensive agent-based simulations to investigate the performance of a system of N agents in solving a cryptarithmetic problem. Cooperation is taken into account through imitative learning which allows information to pass from one agent to another. At each trial the agents can either perform individual trial-and-test operations to explore the solution space or copy cues from a model agent, i.e., the agent that exhibits the lowest cost solution at the trial. We find a trade-off between the number of trial-and-test operations and the number of imitation attempts: too much imitation results in a performance which is poorer than that exhibited by noncooperative agents. For the optimal balance between trial-and-test operations and imitation attempts we find a thirtyfold speedup of the mean time to find the correct solution with respect to the time taken by the noncooperative group. Most significantly, we find that increasing the number of agents N beyond a certain value can greatly harm the performance of the cooperative system which can then perform much worse than in the noncooperative case. Low diversity and the following of a bad leader are the culprits for the poor performance in this case.
The young math learner must make the transition from a concrete number situation, such as that of counting objects (fingers often being the most readily available), to that of using a written symbolic form that stands for the quantities the sets of objects come to represent. This challenging process is often coupled to that of learning a verbal number system that is not always transparent to children. A number of theoretical approaches have been advanced to explain aspects of how this transition takes place in cognitive development. The results obtained with the model presented here, show that a symbol grounding approach can be used to implement aspects of this transition in a cognitive robot. In the current extended version, the model develops finger and word representations, through the use of finger counting and verbal counting strategies, together with the visual representations of learned number symbols, which it uses to perform basic arithmetic operations. In the final training phases, the model is able to do this using only the number symbols as addends. We consider this an example of symbolic grounding, in that through the direct sensory experience with the body (finger counting), a category of linguistic symbol is learned (number words), and both types of representations subsequently serve to ground higher level (numerical) symbols, which are later used exclusively to perform the arithmetic operations.
In this paper, we report the results of a study on the post-intervention effects of applying anodal transcranial direct current stimulation (A-tDCS) on the intensity of P300 potentials. Each of the eight subjects were given both 15 minutes sham and 1.5 mA tDCS in randomized order, in two separate experiments separated by 1 week. The interventions were double blinded. Post-intervention EEG was then recorded after each experiment while subjects were asked to perform a spelling task based on the "odd ball paradigm". Results show a 22% difference, in normalized signal power between tDCS and sham when recorded at 250ms-450ms with a paired t-test p value of 0.057.
The paper presents a control approach based on vertebrate neuromodulation and its implementation on autonomous robots in the open-source, open-access environment of robot operating system (ROS) within a cloud computing framework. A spiking neural network (SNN) is used to model the neuromodulatory function for generating context based behavioral responses of the robots to sensory input signals. The neural network incorporates three types of neurons- cholinergic and noradrenergic (ACh/NE) neurons for attention focusing and action selection, dopaminergic (DA) neurons for rewards- and curiosity-seeking, and serotonergic (5-HT) neurons for risk aversion behaviors. The model depicts description of neuron activity that is biologically realistic but computationally efficient to allow for large-scale simulation of thousands of neurons. The model is implemented using graphics processing units (GPUs) for parallel computing in real-time using the ROS environment. The model is implemented to study the risk-taking, risk-aversive, and distracted behaviors of the neuromodulated robots in single- and multi-robot configurations. The entire process is implemented in a distributed computing framework using ROS where the robots communicate wirelessly with the computing nodes through the on-board laptops. Results are presented for both single- and multi-robot configurations demonstrating interesting behaviors.
Cognition is the product of activation of billions of neurons and their timely interactions. While the activity of individual neurons is essential for proper functioning of the brain, the communication among them is arguably more vital. Previous studies of brain connectivity have largely focused on investigating causality across the brain in order to reveal the existing communication channels that form its internal networks. However, little is known about how these neuronal pathways respond to task demands with varying degrees of complexity. Towards understanding the pathways of information flow, we investigated the effect of memory load on network connectivity of brain. Independent component analysis (ICA) was used to identify brain areas, active during a working memory task, whose activations co-varied with memory load. An information theoretic metric called transfer entropy was adopted to examine the directed links across these areas. Empirical results suggest that the information flow rate across a primary working memory network is modulated by memory load. Furthermore, it was observed that the information flow is affected in pathways with opposite direction during encoding and maintenance stages of working memory operation.
We develop an arousal-based neural model of infant attachment using a deep learning architecture. We show how our model can differentiate between attachment classifications during strange situation-like separation and reunion episodes, in terms of both signalling behaviour and patterns of autonomic arousal, according to the sensitivity of previous interaction.
