A significant challenge in developing translation systems for the world’s ∼7,000 languages is that very few have sufficient data for state-of-the-art techniques. Transfer learning is a promising direction for low-resource neural machine translation (NMT), but introduces many new variables which are often selected through ablation studies, costly trial-and-error, or niche expertise. When pre-training an NMT system for low-resource translation, the pre-training task is often chosen based on data abundance and similarity to the main task. Factors such as dataset sizes and similarity have typically been analysed independently in previous studies, due to the computational cost associated with systematic studies. However, these factors are not independent. We conducted a three-factor experiment to examine how language similarity, pre-training dataset size and main dataset size interacted in their effect on performance in pre-trained transformer-based low-resource NMT. We replicated the common finding that more data was beneficial in bilingual systems, but also found a statistically significant interaction between the three factors, which reduced the effectiveness of large pre-training datasets for some main task dataset sizes (p-value < 0.0018). The surprising trends identified in these interactions indicate that systematic studies of interactions may be a promising long-term direction for guiding research in low-resource neural methods.

Noise and temporal dynamics are ubiquitous in neural systems yet the computational consequences of these two phenomena interacting are not well studied. Temporal dynamics in spiking networks are often considered only implicitly as part of membrane time constants or synaptic transfer functions. We explicitly model temporal structure using plastic conduction delays between neuron's and characterise the influence of different kinds of noise on learning including temporal jitter, dropout, pattern size, and pattern presentation frequency. We simplify the conduction delay plasticity (CDP) rule called synaptic delay variance learning (SDVL) and demonstrate it is robust to several kinds of noise including; internal pattern jitter, number of pattern afferents, pattern presentation rate, and pattern spike dropout. In particular, after unsupervised training the simplified version of SDVL can achieve an accuracy of up to 99.7 percent averaged over 100 trials. These results demonstrate that learning algorithms based on explicitly modelling temporal structure in inputs can be functional and robust for unsupervised learning of spatiotemporal patterns across a range of noise conditions.

Calculating the Shannon entropy for symbolic sequences has been widely considered in many fields. For descriptive statistical problems such as estimating the N-gram entropy of English language text, a common approach is to use as much data as possible to obtain progressively more accurate estimates. However, in some instances, only short sequences may be available. This gives rise to the question of how many samples are needed to compute entropy. In this paper, we examine this problem and propose a method for estimating the number of samples required to compute Shannon entropy for a set of ranked symbolic “natural” events. The result is developed using a modified Zipf-Mandelbrot law and the Dvoretzky-Kiefer-Wolfowitz inequality, and we propose an approximation which yields an estimate for the minimum number of samples required to obtain an estimate of entropy with a given confidence level and degree of accuracy.

The use of robots, as a social stimulus, provides several advantages over using another animal. In particular, for rat-robot studies, robots can produce social behaviour that is reproducible across trials. In the current work, we outline a framework for rat-robot interaction studies, that consists of a novel rat-sized robot (PiRat), models of robotic behavior, and a position tracking system for both robot and rat. We present the design of the framework, including constraints on autonomy, latency, and control. We pilot tested our framework by individually running the robot rat with eight different rats, first through a habituation stage, and then with PiRat performing two different types of behaviour - avoiding and frequently approaching. We evaluate the performance of the framework on latency and autonomy, and on the ability to influence the behaviour of individual rats. We find that the framework performs well on its constraints, engages some of the rats (according to the number of meetings), and features a control scheme that produces reproducible behaviour in rats. These features represent a first demonstration of a closed-loop rat-robot framework.

Motion is a critical aspect of communication, required to create natural interactions between humans and robots. Robots for the classroom pose several constraints on motion, which make them challenging to design, including maintaining the safety of the child and the robot, responding in a timely fashion, and creating motions that are expressive and not scary. In this paper we present the mechanical design of a social robot and demonstrate that it is capable of safe motion within the proximity of children through analysis and empirical testing of the arms. The robot has a novel mechanical design for its two arms, which include torso-mounted, back-drivable, torque-limited stepper motors. The results suggest that our design succeeds at increasing safety levels while enabling the use of socially acceptable speeds of motion during the interaction. This study implies that the design of robotic agents for social interaction with children should consider the design of mechanical features that enable safe contact between the human and the robot while not limiting the robot to slow motions that would impair the timing of the interaction.

Language can be a useful tool for social robots as part of their repertoire of social engagement. This late breaking report outlines preliminary studies into how a child can teach a robot lexicons for colors and color relations. The robot used is a minimal social robot, made from cardboard and foam, that interacts with the children through a simple color naming game. Distributed, non-parametric lexicons similar to those used in previous language learning robot studies are used to store links between words and colors. We visually present the resulting lexicons and highlight the issues that have arisen from this preliminary study and how they can be resolved for future studies. The results of this study indicate that children can teach a social robot lexicons, allowing the children and robot to develop a shared set of symbols for color.

Robots that facilitate touch by children have special requirements in terms of safety and robustness, but little is known about how and when children actually use touch with robots. Tools and techniques are required to sense the variety of children's touch and to interpret the volumes of data generated. This explorative user study investigated children's patterns of touch during game play with a robot. We examined where the children touch the robot and their patterns of touch over time, using a raster-based visualisation of each child's time series of touches, recording patterns of touch across different games and children. We found that children readily engage with the robot, in particular spontaneously touching the robot's hands more than any other area. This user study and the tools developed may aid future designs of robots to autonomously detect when they have been touched.

An important challenge in Human-Robot Interaction (HRI) is the analysis and interpretation of large volumes of data to inform the design process within a complex, multi-faceted research space. In this study, we explore how data visualization techniques can contribute to HRI methodology, particularly in terms of linking qualitative and quantitative analysis methods. Specifically, we present a case study demonstrating the visualization of touch data to identify potential patterns of interaction with a social robot intended for deployment in preschool classrooms.

Pretotyping is a set of techniques, tools, and metrics for gauging the interest in a product, prior to full-scale development [1]. This late breaking report describes a pretotyping case study of an ethnodroid - a robot that functions as an ethnographer - intended to engage with young children and record their learning progress. The central requirement for the project is that the robot will be able to interact socially with children aged 1-6 years in tablet-based tasks. We developed a simple robot made of MDF (thick cardboard), added tablets for the face and torso, and controlled a scripted interaction using Wizard of Oz (WoZ). Children's engagement with the robot was tested in an early learning centre which provided a relatively structured environment (“in the lab”) and at a science fair which provided a relatively unconstrained setting (“in the wild”). The rapid testing revealed distinct effects in the children's attitudes and behaviors in the two user contexts and provided insights into form, sensors and analyses for the design process.

While information visualization frameworks and heuristics have traditionally been reluctant to include acquired codes of meaning, designers are making use of them in a wide variety of ways. Acquired codes leverage a user's experience to understand the meaning of a visualization. They range from figurative visualizations which rely on the reader's recognition of shapes, to conventional arrangements of graphic elements which represent particular subjects. In this study, we used content analysis to codify acquired meaning in visualization. We applied the content analysis to a set of infographics and data visualizations which are exemplars of innovative and effective design. 88% of the infographics and 71% of data visualizations in the sample contain at least one use of figurative visualization. Conventions on the arrangement of graphics are also widespread in the sample. In particular, a comparison of representations of time and other quantitative data showed that conventions can be specific to a subject. These results suggest that there is a need for information visualization research to expand its scope beyond perceptual channels, to include social and culturally constructed meaning. Our paper demonstrates a viable method for identifying figurative techniques and graphic conventions and integrating them into heuristics for visualization design.

This paper presents a novel paradigm for a spiking neural network to forecast temporal sequences. The key to the approach is a new model of a spiking neuron that can make multi-step predictions, using learnable temporal delays at both dendrites and axons. This model is able to learn the temporal structure of space-time events, adaptable to multiple scales, with the neurons able to function asynchronously to predict future events in a video sequence. This approach contrasts with conventional neural network approaches that use fixed time steps and iterative prediction. Simulations were conducted to compare the new model to a conventional iterative paradigm on motion sequences from a frame-free event-driven Dynamic Vision Sensor (DVS128, 16k pixels), showing that the new approach consistently has a low prediction error while the iterative paradigm is affected by propagated errors.

BACKGROUND: Erroneous patient birthdates are common in health databases. Detection of these errors usually involves manual verification, which can be resource intensive and impractical. By identifying a frequent manifestation of birthdate errors, this paper presents a principled and statistically driven procedure to identify erroneous patient birthdates. RESULTS: Generalized additive models (GAM) enabled explicit incorporation of known demographic trends and birth patterns. With false positive rates controlled, the method identified birthdate contamination with high accuracy. In the health data set used, of the 58 actual incorrect birthdates manually identified by the domain expert, the GAM-based method identified 51, with 8 false positives (resulting in a positive predictive value of 86.0% (51/59) and a false negative rate of 12.0% (7/58)). These results outperformed linear time-series models. CONCLUSIONS: The GAM-based method is an effective approach to identify systemic birthdate errors, a common data quality issue in both clinical and administrative databases, with high accuracy.

Previous studies have shown how Lingodroids, language learning mobile robots, learn terms for space and time, connecting their personal maps of the world to a publically shared language. One caveat of previous studies was that the robots shared the same cognitive architecture, identical in all respects from sensors to mapping systems. In this paper we investigate the question of how terms for space can be developed between robots that have fundamentally different sensors and spatial representations. In the real world, communication needs to occur between agents that have different embodiment and cognitive capabilities, including different sensors, different representations of the world, and different species (including humans). The novel aspects of these studies is that one robot uses a forward facing camera to estimate appearance and uses a biologically inspired continuous attractor network to generate a topological map; the other robot uses a laser scanner to estimate range and uses a probabilistic filter approach to generate an occupancy grid. The robots hold conversations in different locations to establish a shared language. Despite their different ways of sensing and mapping the world, the robots are able to create coherent lexicons for the space around them.

For humans and robots to communicate using natural language it is necessary for the robots to develop concepts and associated terms that correspond to the human use of words. Time and space are foundational concepts in human language, and to develop a set of words that correspond to human notions of time and space, it is necessary to take into account the way that they are used in natural human conversations, where terms and phrases such as `soon', `in a while', or `near' are often used. We present language learning robots called Lingodroids that can learn and use simple terms for time and space. In previous work, the Lingodroids were able to learn terms for space. In this work we extend their abilities by adding temporal variables which allow them to learn terms for time. The robots build their own maps of the world and interact socially to form a shared lexicon for location and duration terms. The robots successfully use the shared lexicons to communicate places and times to meet again.

Transmission delays are an inherent component of spiking neural networks (SNNs) but relatively little is known about how delays are adapted in biological systems and studies on computational learning mechanisms have focused on spike-timing-dependent plasticity (STDP) which adjusts synaptic weights rather than synaptic delays. We propose a novel algorithm for learning temporal delays in SNNs with Gaussian synapses, which we call spike-delay-variance learning (SDVL). A key feature of the algorithm is adaptation of the shape (mean and variance) of the postsynaptic release profiles only, rather than the conventional STDP approach of adapting the network's synaptic weights. The algorithm's ability to learn temporal input sequences was tested in three studies using supervised and unsupervised learning within feed-forward networks. SDVL was able to successfully classify forty spatiotemporal patterns without supervision by providing robust, effective adaption of the postsynaptic release profiles. The results demonstrate how delay learning can contribute to the stability of spiking sequences, and that there is a potential role for adaption of variance as well as mean values in learning algorithms for spiking neural networks.

In this paper we present the design and implementation of a low power neural amplifier which has two programmable gains in two programmable bandwidths. The bandwidths are programmable between 0.7-300Hz, suitable for measuring local field potentials and 1.95-5.4kHz, suitable for measuring action potentials. The amplifier achieves a maximum gain of 79dB in the higher bandwidth. A chip has been designed and implemented using a 0.5μm technology with 8 neural amplifiers. On average the neural amplifier consumes less than 14μW at 3.3V.

Spatial navigation requires the processing of complex, disparate and often ambiguous sensory data. The neurocomputations underpinning this vital ability remain poorly understood. Controversy remains as to whether multimodal sensory information must be combined into a unified representation, consistent with Tolman's "cognitive map", or whether differential activation of independent navigation modules suffice to explain observed navigation behaviour. Here we demonstrate that key neural correlates of spatial navigation in darkness cannot be explained if the path integration system acted independently of boundary (landmark) information. In vivo recordings demonstrate that the rodent head direction (HD) system becomes unstable within three minutes without vision. In contrast, rodents maintain stable place fields and grid fields for over half an hour without vision. Using a simple HD error model, we show analytically that idiothetic path integration (iPI) alone cannot be used to maintain any stable place representation beyond two to three minutes. We then use a measure of place stability based on information theoretic principles to prove that featureless boundaries alone cannot be used to improve localization above chance level. Having shown that neither iPI nor boundaries alone are sufficient, we then address the question of whether their combination is sufficient and--we conjecture--necessary to maintain place stability for prolonged periods without vision. We addressed this question in simulations and robot experiments using a navigation model comprising of a particle filter and boundary map. The model replicates published experimental results on place field and grid field stability without vision, and makes testable predictions including place field splitting and grid field rescaling if the true arena geometry differs from the acquired boundary map. We discuss our findings in light of current theories of animal navigation and neuronal computation, and elaborate on their implications and significance for the design, analysis and interpretation of experiments.

Human communication is more than just the transmission of information. It also involves complex interaction dynamics that reflect the roles and communication styles of the participants. A novel approach to studying human communication is to view conversation as a coupled time series and apply analysis techniques from dynamical systems to the recurring topics or concepts. In this paper, we define a set of metrics that enable quantification of the complex interaction dynamics visible in conceptual recurrence. These multi-participant recurrence (MPR) metrics can be seen as an extension of recurrence quantification analysis (RQA) into the symbolic domain. This technique can be used to monitor the state of a communication system and inform about interaction dynamics, including the level of topic consistency between participants; the timing of state changes for the participants as a result of changes in topic focus; and, patterns of topic proposal, reflection, and repetition. We demonstrate three use studies applying the new metrics to conversation transcripts from different genres to demonstrate their ability to characterize individual communication participants and intergroup communication patterns.

Human discourse contains a rich mixture of conceptual information. Visualization of the global and local patterns within this data stream is a complex and challenging problem. Recurrence plots are an information visualization technique that can reveal trends and features in complex time series data. The recurrence plot technique works by measuring the similarity of points in a time series to all other points in the same time series and plotting the results in two dimensions. Previous studies have applied recurrence plotting techniques to textual data; however, these approaches plot recurrence using term-based similarity rather than conceptual similarity of the text. We introduce conceptual recurrence plots, which use a model of language to measure similarity between pairs of text utterances, and the similarity of all utterances is measured and displayed. In this paper, we explore how the descriptive power of the recurrence plotting technique can be used to discover patterns of interaction across a series of conversation transcripts. The results suggest that the conceptual recurrence plotting technique is a useful tool for exploring the structure of human discourse.

The Lingodroids are a pair of mobile robots that evolve a language for places and relationships between places (based on distance and direction). Each robot in these studies has its own understanding of the layout of the world, based on its unique experiences and exploration of the environment. Despite having different internal representations of the world, the robots are able to develop a common lexicon for places, and then use simple sentences to explain and understand relationships between places even places that they could not physically experience, such as areas behind closed doors. By learning the language, the robots are able to develop representations for places that are inaccessible to them, and later, when the doors are opened, use those representations to perform goal-directed behavior.

To successfully navigate their habitats, many mammals use a combination of two mechanisms, path integration and calibration using landmarks, which together enable them to estimate their location and orientation, or pose. In large natural environments, both these mechanisms are characterized by uncertainty: the path integration process is subject to the accumulation of error, while landmark calibration is limited by perceptual ambiguity. It remains unclear how animals form coherent spatial representations in the presence of such uncertainty. Navigation research using robots has determined that uncertainty can be effectively addressed by maintaining multiple probabilistic estimates of a robot's pose. Here we show how conjunctive grid cells in dorsocaudal medial entorhinal cortex (dMEC) may maintain multiple estimates of pose using a brain-based robot navigation system known as RatSLAM. Based both on rodent spatially-responsive cells and functional engineering principles, the cells at the core of the RatSLAM computational model have similar characteristics to rodent grid cells, which we demonstrate by replicating the seminal Moser experiments. We apply the RatSLAM model to a new experimental paradigm designed to examine the responses of a robot or animal in the presence of perceptual ambiguity. Our computational approach enables us to observe short-term population coding of multiple location hypotheses, a phenomenon which would not be easily observable in rodent recordings. We present behavioral and neural evidence demonstrating that the conjunctive grid cells maintain and propagate multiple estimates of pose, enabling the correct pose estimate to be resolved over time even without uniquely identifying cues. While recent research has focused on the grid-like firing characteristics, accuracy and representational capacity of grid cells, our results identify a possible critical and unique role for conjunctive grid cells in filtering sensory uncertainty. We anticipate our study to be a starting point for animal experiments that test navigation in perceptually ambiguous environments.

Concepts maps are a useful method for visualizing how concepts are associated within a domain. Automated tools can extract the key concepts from a corpus and display their relationships in a two-dimensional representation. However, it may be difficult to quickly perceive the overall structure of a concept map. The aim of this research is to provide a framework on a concept map to identify primary concepts and nearby, related concepts providing context for them using network graph techniques. Automated concept maps are usually very highly connected, and full connectivity shows none of the detail or fine structure. Minimum spanning trees based on the strength of the co-occurrence frequency between concepts and their relative distances on the two-dimensional concept map show hierarchical structure, with key concepts acting as hubs within the network structure. The links within the minimum spanning tree can be used to associate concepts that are not directly connected.

This article describes an interactive visualization tool, LinMap, for exploring the structure of complexity gradients in evolutionary landscapes. LinMap is a computationally efficient and intuitive tool for visualizing and exploring multidimensional parameter spaces. An artificial cell lineage model is presented that allows complexity to be quantified according to several different developmental and phenotypic metrics. LinMap is applied to the evolutionary landscapes generated by this model to demonstrate that different definitions of complexity produce different gradients across the same landscape; that landscapes are characterized by a phase transition between proliferating and quiescent cell lineages where both complexity and diversity are maximized; and that landscapes defined by adaptive fitness and complexity can display different topographical features.

The general assumption in biology is that most genes encode proteins. However, it is now evident that much of the genome of humans and other complex organisms is transcribed into non-protein-coding RNAs. Some of these RNAs are processed into small regulatory RNAs, such as microRNAs, that control many aspects of animal and plant development. It has been suggested that regulatory RNAs represent an additional control layer that was critical to the emergence of complex organisms. We examine this possibility using a model of cell differentiation based on attractors in boolean networks. Our simulation studies show that an additional layer of RNA control modeled as fast temporal links can significantly increase the number of attractors in boolean models of genetic regulatory networks (analogous to the number of cell types in a complex organism). However, it also has the power to simplify the state space structure. We explore the conditions under which these different outcomes occur.

We investigate the effects of constitutive gene activation upon the dynamics of random Boolean networks using a suite of models. We find that constitutive activity leads to simpler state spaces, with fewer and larger basins of attraction. The major difference in patterns of basin number and distribution is seen between networks with no activation and those with a single constitutively active node. Increasing the proportion of constitutively active node intensifies, but does not qualitatively change, the observed patterns. The proportion of genes constitutively active interacts in a nonlinear way with other network parameters, such as average connectivity and proportion of inhibitory links. We conclude that constitutive gene activation has a fundamental effect on network behavior that has been overlooked in previous random Boolean network studies. It acts to constrain the start states of the networks, and the states which can be reached during development and differentiation, and we hypothesize that constitutive activation and repression of genes may help to guide the process whereby a single genetic regulatory network produces a range of different cell types.

Human perception is finely tuned to extract structure about the 4D world of time and space as well as properties such as color and texture. Developing intuitions about spatial structure beyond 4D requires exploiting other perceptual and cognitive abilities. One of the most natural ways to explore complex spaces is for a user to actively navigate through them, using local explorations and global summaries to develop intuitions about structure, and then testing the developing ideas by further exploration. This article provides a brief overview of a technique for visualizing surfaces defined over moderate-dimensional binary spaces, by recursively unfolding them onto a 2D hypergraph. We briefly summarize the uses of a freely available Web-based visualization tool, Hyperspace Graph Paper (HSGP), for exploring fitness landscapes and search algorithms in evolutionary computation. HSGP provides a way for a user to actively explore a landscape, from simple tasks such as mapping the neighborhood structure of different points, to seeing global properties such as the size and distribution of basins of attraction or how different search algorithms interact with landscape structure. It has been most useful for exploring recursive and repetitive landscapes, and its strength is that it allows intuitions to be developed through active navigation by the user, and exploits the visual system's ability to detect pattern and texture. The technique is most effective when applied to continuous functions over Boolean variables using 4 to 16 dimensions.

NK models provide a family of tunably rugged fitness landscapes used in a wide range of evolutionary computation studies. It is well known that the average height of local optima regresses to the mean of the landscape with increasing epistasis, k. This fact has been confirmed using both theoretical studies of landscape structure and empirical studies of evolutionary search. We show that the global optimum behaves quite differently: the expected value of the global maximum is highest in the maximally rugged case. Furthermore, we demonstrate that this expected value increases with K, despite the fact that the average fitness of the local optima decreases. That is, the highest peaks are found in the most rugged landscapes, scattered amongst masses of low-lying peaks. We find the asymptotic value of the global optimum as N approaches infinity for both the smooth and maximally rugged cases. In evolutionary search, the optima that are found reflect the local optima that exist in the landscape, the size of these optima -- which corresponds to the size of their basins of attraction, and the effort expended in the search process. Increasing the level of epistasis in an NK landscape stochastically introduces higher peaks, but renders them exponentially more difficult to find.

Biological development is a remarkably complex process. A single cell, in an appropriate environment, contains sufficient information to generate a variety of differentiated cell types, whose spatial and temporal dynamics interact to form detailed morphological patterns. While several different physical and chemical processes play an important role in the development of an organism, the locus of control is the cell's gene regulatory network. We designed a dynamic recurrent gene network (DRGN) model and evaluated its ability to control the developmental trajectories of cells during embryogenesis. Three tasks were developed to evaluate the model, inspired by cell lineage specification in C. elegans, describing the variation in gene activity required for early cell diversification, combinatorial control of cell lineages, and cell lineage termination. Three corresponding sets of simulations compared performance on the tasks for different gene network sizes, demonstrating the ability of DRGNs to perform the tasks with minimal external input. The model and task definition represent a new means of linking the fundamental properties of genetic networks with the topology of the cell lineages whose development they control.

Deacon has recently proposed that complexes of genes can be integrated into functional groups as a result of environmental changes that mask and unmask selection pressures. For example, many animals endogenously synthesize ascorbic acid (vitamin C), but anthropoid primates have only a nonfunctional version of the crucial gene for this pathway. It is hypothesized that the loss of functionality occurred in the evolutionary past when a diet rich in vitamin C masked the effect of the gene, and its loss effectively trapped the animals in a fruit-eating lifestyle. As a result, the complex of abilities that support this lifestyle were evolutionarily bound together, forming a multilocus complex. In this study we use evolutionary computation simulations to explore the thesis that masking and unmasking can transfer dependence from one set of genes to many sets, and thereby integrate the whole complex of genes. We used a framework based on Hinton and Nowlan's 1987 simulation of the Baldwin effect. Additional gene complexes and an environmental parameter were added to their basic model, and the fitness function extended. The simulation clearly demonstrates that the genetic redistribution effect can occur in silico, showing an initial advantage of endogenously synthesized vitamin C, followed by transfer of the fitness contribution to the complex of genes that together allow the acquisition of vitamin C from the environment. As is well known in the modeling community, the Baldwin effect only occurs in simulations when the population of agents is ''poised on the brink'' of discovering the genetically specified solution. Similarly, the redistribution effect occurs in simulations under specific initial conditions: too little vitamin C in the environment, and its synthesis it is never fully masked; too much vitamin C, and the abilities required to acquire it are not tightly integrated. The Baldwin effect has been hypothesized as a potential mechanism for developing language-specific adaptations like innate universal grammar and other highly modular capacities. We conclude with a discussion of the relevance of genetic assimilation and genetic redistribution to the evolution of language and other cognitive adaptations.

As advances in molecular biology continue to reveal additional layers of complexity in gene regulation, computational models need to incorporate additional features to explore the implications of new theories and hypotheses. It has recently been suggested that eukaryotic organisms owe their phenotypic complexity and diversity to the exploitation of small RNAs as signalling molecules. Previous models of genetic systems are, for several reasons, inadequate to investigate this theory. In this study, we present an artificial genome model of genetic regulatory networks based upon previous work by Torsten Reil, and demonstrate how this model generates networks with biologically plausible structural and dynamic properties. We also extend the model to explore the implications of incorporating regulation by small RNA molecules in a gene network. We demonstrate how, using these signals, highly connected networks can display dynamics that are more stable than expected given their level of connectivity.

Complex systems techniques provide a powerful tool to study the emergent properties of networks of interacting genes. In this study we extract models of genetic regulatory networks from an artificial genome, represented by a sequence of nucleotides, and analyse how variations in the connectivity and degree of inhibition of the extracted networks affects the resulting classes of behaviours. For low connectivity systems were found to be very stable. Only with higher connectivity was a significant occurrence of chaos found. Most interestingly, the peak in occurrence of chaos occurs perched on the edge of a phase transition in the occurrence of null attractors.

In this paper we present a technique for visualising hierarchical and symmetric, multi-modal fitness functions that have been investigated in the evolutionary computation literature. The focus of this technique is on landscapes in moderate-dimensional, binary spaces (i.e., fitness functions defined over [0,1](n), for n < or = 16). The visualisation approach involves an unfolding of the hyperspace into a two-dimensional graph, whose layout represents the topology of the space using a recursive relationship, and whose shading defines the shape of the cost surface defined on the space. Using this technique we present case-study explorations of three fitness functions: royal road, hierarchical-if-and-only-if (H-IFF), and hierarchically decomposable functions (HDF). The visualisation approach provides an insight into the properties of these functions, particularly with respect to the size and shape of the basins of attraction around each of the local optima.

It has been demonstrated that several standard evolutionary computation test problems can be solved by a simple hill climbing search algorithm-often more efficiently than by a population based evolutionary algorithm. There remain some classes of problems, however, for which maintaining a genetically diverse population is essential in order to discover the optimal solution. In biological populations, diversity maintenance is important to enable populations to adapt to rapidly changing environments and to exploit environmental niches. We demonstrate that on a neutral landscape recombination allows a population to maintain a significantly greater level of genetic diversity through the transition between two fitness layers. Recombination may therefore have a role to play in maintaining population diversity across fitness transitions.

Recent research in molecular evolution has raised awareness of the importance of selective neutrality. Several different models of neutrality have been proposed based on Kauffman's well-known NK landscape model. Two of these models, NKp and NKq, are investigated and found to display significantly different structural properties. The fitness distributions of these neutral landscapes reveal that their levels of correlation with non-neutral landscapes are significantly different, as are the distributions of neutral mutations. In this paper we describe a series of simulations of a hill climbing search algorithm on NK, NKp and NKq landscapes with varying levels of epistatic interaction. These simulations demonstrate differences in the way that epistatic interaction affects the 'searchability' of neutral landscapes. We conclude that the method used to implement neutrality has an impact on both the structure of the resulting landscapes and on the performance of evolutionary search algorithms on these landscapes. These model-dependent effects must be taken into consideration when modelling biological phenomena.

The genetic assimilation of learned behaviour was introduced to a wider evolutionary computation field by the classic simulation of Hinton and Nowlan (1987). Subsequent studies have analysed and extended their initial framework, contributing to the understanding of the often counterintuitive relationship between evolution and learning. We add to this increasing body of literature by presenting an evolving population of neural networks that plainly exhibit the Baldwin effect. Phenotypic plasticity, embodied in the literal learning rate of the neural networks, is evolved along with the network connection weights. Significantly, high levels of plasticity do not cause the population to genetically stagnate once correct behaviour can be learned. Rather, continuing inter-population competition drives the levels of learning down as beneficial behaviour becomes genetically specified. By observing the evolving learning rate of the agent population, and by comparing the learned and innate agent responses, we demonstrate the Baldwin effect in its entirety.

We present a technique for visualising cost surfaces that are relevant to work in evolutionary computation, particularly genetic algorithms. The focus is on moderate-dimensional, binary cost surfaces (i.e., functions defined over {0, 1}/sup n/, for n/spl les/16) that have a hierarchical, modular structure. The visualisation approach involves an unfolding of the hyperspace into a two-dimensional graph, whose layout represents the topology of the space using a recursive relationship, and whose shading defines the shape of the cost surface defined on the space. Using this technique we present a case-study exploration of the hierarchical-if-and-only-if (H-IFF) function. The visualisation approach provides an insight into the properties of this function, particularly in regards to the size and shape of the basins of attraction around each of the local optima.

The original demonstration by G. Hinton and S. Nowlan (1987) of the Baldwin effect (J. Baldwin, 1896) is well-known and serves as an interesting basis for genetic algorithm (GA) research. A variant of the original representation used a binary code, in which learning was expressed as a substitute for internalised knowledge; in this paper, the representation is altered such that learning becomes an expression of uncertainty. This change results in an interesting and non-trivial set of interactions between the GA operators and the representation, as well as enhancing the performance and robustness of the GA.

The long short-term memory (LSTM) is not the only neural network which learns a context sensitive language. Second-order sequential cascaded networks (SCNs) are able to induce means from a finite fragment of a context-sensitive language for processing strings outside the training set. The dynamical behavior of the SCN is qualitatively distinct from that observed in LSTM networks. Differences in performance and dynamics are discussed.

Artificial intelligence (AI) is playing an increasingly important role in the success or failure of computer games. The paper explores the use of AI in games focusing on difficulties faced by game developers and techniques that have proven useful. An argument is presented for the value of collaboration between the game and academic AI communities.

This paper describes a stochastic implementation of Austin et al.'s (1999) model of the spread of antibiotic resistance in a population of fixed size under varying conditions of antibiotic use. The population is divided into sub-groups: individuals colonized by commensal bacteria and an uncolonized group. The colonized group is further divided according to whether the commensal bacteria are sensitive or resistant to antibiotics. This study uses Monte Carlo techniques to model the dynamics of the evolution of the antibiotic resistant population, a study that cannot be done in the original model. The Monte Carlo approach allows the investigation of the transient dynamics of the spread of resistance, the effects of finite (especially small) populations and the interaction of model parameters.

Continuous-valued recurrent neural networks can learn mechanisms for processing context-free languages. The dynamics of such networks is usually based on damped oscillation around fixed points in state space and requires that the dynamical components are arranged in certain ways. It is shown that qualitatively similar dynamics with similar constraints hold for anbncn , a context-sensitive language. The additional difficulty with anbncn , compared with the context-free language anbn , consists of 'counting up' and 'counting down' letters simultaneously. The network solution is to oscillate in two principal dimensions, one for counting up and one for counting down. This study focuses on the dynamics employed by the sequential cascaded network, in contrast to the simple recurrent network, and the use of backpropagation through time. Found solutions generalize well beyond training data, however, learning is not reliable. The contribution of this study lies in demonstrating how the dynamics in recurrent neural networks that process context-free languages can also be employed in processing some context-sensitive languages (traditionally thought of as requiring additional computation resources). This continuity of mechanism between language classes contributes to our understanding of neural networks in modelling language learning and processing.

August 01 1996 Using Bottlenecks in Feedforward Networks as a Dimension Reduction Technique: An Application to Optimization Tasks In Special Collection: CogNet Janet Wiles, Janet Wiles Departments of Computer Science and Psychology, The University of Queensland, Queensland 4072, Australia Search for other works by this author on: This Site Google Scholar Paul Bakker, Paul Bakker Departments of Computer Science and Psychology, The University of Queensland, Queensland 4072, Australia Search for other works by this author on: This Site Google Scholar Adam Lynton, Adam Lynton Department of Computer Science, The University of Queensland, Queensland 4072, Australia Search for other works by this author on: This Site Google Scholar Michael Norris, Michael Norris Department of Computer Science, The University of Queensland, Queensland 4072, Australia Search for other works by this author on: This Site Google Scholar Sean Parkinson, Sean Parkinson Department of Computer Science, The University of Queensland, Queensland 4072, Australia Search for other works by this author on: This Site Google Scholar Mark Staples, Mark Staples Department of Computer Science, The University of Queensland, Queensland 4072, Australia Search for other works by this author on: This Site Google Scholar Alan Whiteside Alan Whiteside Department of Computer Science, The University of Queensland, Queensland 4072, Australia Search for other works by this author on: This Site Google Scholar Author and Article Information Janet Wiles Departments of Computer Science and Psychology, The University of Queensland, Queensland 4072, Australia Paul Bakker Departments of Computer Science and Psychology, The University of Queensland, Queensland 4072, Australia Adam Lynton Department of Computer Science, The University of Queensland, Queensland 4072, Australia Michael Norris Department of Computer Science, The University of Queensland, Queensland 4072, Australia Sean Parkinson Department of Computer Science, The University of Queensland, Queensland 4072, Australia Mark Staples Department of Computer Science, The University of Queensland, Queensland 4072, Australia Alan Whiteside Department of Computer Science, The University of Queensland, Queensland 4072, Australia Received: December 13 1993 Accepted: January 22 1996 Online ISSN: 1530-888X Print ISSN: 0899-7667 © 1996 Massachusetts Institute of Technology1996 Neural Computation (1996) 8 (6): 1179–1183. https://doi.org/10.1162/neco.1996.8.6.1179 Article history Received: December 13 1993 Accepted: January 22 1996 Cite Icon Cite Permissions Share Icon Share Facebook Twitter LinkedIn Email Views Icon Views Article contents Figures & tables Video Audio Supplementary Data Peer Review Search Site Citation Janet Wiles, Paul Bakker, Adam Lynton, Michael Norris, Sean Parkinson, Mark Staples, Alan Whiteside; Using Bottlenecks in Feedforward Networks as a Dimension Reduction Technique: An Application to Optimization Tasks. Neural Comput 1996; 8 (6): 1179–1183. doi: https://doi.org/10.1162/neco.1996.8.6.1179 Download citation file: Ris (Zotero) Reference Manager EasyBib Bookends Mendeley Papers EndNote RefWorks BibTex toolbar search Search Dropdown Menu toolbar search search input Search input auto suggest filter your search All ContentAll JournalsNeural Computation Search Advanced Search © 1996 Massachusetts Institute of Technology1996 Article PDF first page preview Close Modal You do not currently have access to this content.