ABSTRACT. A driver for achieving human-level AI and high-fidelity cognitive architectures is the ability to easily test and compare the performance and behavior of computational agents/models to humans and to one another. One major difficulty in setting up and getting participation in large-scale Cognitive Decathlon and Grand Challenge competitions, or even smaller scale cross-framework evaluation and Turing testing, is that there is no standard interface protocol that enables and facilitates human and computational agent "plug-and-play" participation across various tasks. We identify three major issues. First, human-readable task interfaces aren't often translated into machine-readable form. Second, in the cases where a task interface is made available in a machine-readable protocol, the protocol is often task-specific, and differs from other task protocols. Finally, where both human and machine-readable versions of the task interface exist, the two versions often differ in content. This makes the bar of entry extremely high for comparison of humans and multiple computational frameworks across multiple tasks. This paper proposes a standard approach to task design where all task interactions adhere to a standard API. We provide examples of how this method can be employed to gather human and computational simulation data in text-and-button tasks, visual and animated tasks, and in real-time robotics tasks.

ABSTRACT. Architectures for neuromorphic computing have evolved to the point where a significant amount of software design and implementation is necessary for leveraging these architectures to solve real problems. In this paper, we describe the software ecosystem for the DANNA neuromorphic computing model. This ecosystem is composed of four pieces: a simulator, commander, evolutionary optimizer, and visualizer. The ecosystem facilitates developing applications for DANNA, and in turn will help DANNA evolve into a more effective neuromorphic computing model. We describe how we use the software components to perform digit classification using the MNIST dataset.

ABSTRACT. How do brains learn which features matter how much, when and for what purposes? A specific feature may matter more or less for recognitions of different learned patterns, and in different contexts and attentional foci. Simple "neural circuits" built from biologically-inspired reusable memory pattern components in the NeurOS and NeuroBlocks technology model a range of learning and dynamic contextual/situational/attentional feature relevance. A pattern is a collection of weighted features, roughly analogous to a neuron or neuron assembly. New patterns are created for sufficiently novel feature combinations. Individual feature weights in best-matching existing patterns grow or diminish with repetition, yielding patterns that adjust to repeated experience. Arbitrarily complex classification meshes typical of human knowledge are easily assembled by varying a simple novelty parameter. Cascading pattern recognitions build up layers of concrete to abstract feature vocabularies. Names or labels are modeled as synonyms for experience patterns. Context can be modeled as yet another feature, derived from recent activity, to discriminate among otherwise similar patterns. Attention can be modeled as broad dynamic parameters modulating feature signal strengths.

ABSTRACT. A virtual character (VC) acts within its virtual world boundaries, but with vision sensory capabilities, it may be expected to explore the real world and interact with the intelligent beings there. Such a VC can be equipped with algorithms to localize humans, recognize and communicate with them. These perceptual capabilities prompt a sophisticated cognitive architecture (CA), enabling our VC to learn from intelligent beings and perhaps reason like one. This CA needs to be fairly seamless, reliable and adaptive. Here we explore a vision-based human-centric approach to the VC design.

ABSTRACT. Behavior planning is known to be one of the basic cognitive functions, which is essential for any cognitive architecture of any control system used in robotics. At the same time most of the widespread planning algorithms employed in those systems are developed using only approaches and models of Artificial Intelligence and don’t take into account numerous results of cognitive experiments. As a result, there is a strong need for novel methods of behavior planning suitable for modern cognitive architectures aimed at robot control. One such method is presented in this work and is studied within a special class of navigation task called smart relocation task. The method is based on the hierarchical two-level model of abstraction and knowledge representation, e.g. symbolic and subsymbolic. On the symbolic level sign world model is used for knowledge representation and hierarchical planning algorithm, PMA, is utilized for planning. On the subsymbolic level the task of path planning is considered and solved as a graph search problem. Interaction between both planners is examined and inter-level interfaces and feedback loops are described. Preliminary experimental results are presented.

ABSTRACT. In this paper, we introduce the Cognitive Systems Toolkit (CST) and its underlying CST Cognitive Architecture. CST is a general toolkit for the construction of cognitive architectures, which relies on a set of concepts which are traversal to many other cognitive architecture and constitute CST's core. This core is general enough such that problem-specific cognitive architectures might be generated by using CST. At the same time, CST specify a more general Cognitive Architecture, which is a reference model for the construction of application specific cognitive architectures. Differently from other Cognitive Architectures available in the literature, which are computational frameworks, CST is a toolkit, which means more flexibility in choosing specific techniques and algorithms for setting up the application architecture. After presenting the architecture core, we develop an illustrative example showing how CST can be used to implement a simple subsumption architecture in a robotic application.

ABSTRACT. This work proposes a motivational system for an autonomous robot that guides the fulfilment of its goals in a developmental manner, discovering sub-goals not only as a way to simplify goal achievement, but as a way to acquire knowledge in an incremental, modular and reusable fashion. This system has been called MotivEn (Motivational Engine) and we have performed an initial integration of it with the Multilevel Darwinist Brain (MDB) cognitive architecture. We describe here the main elements of MotivEn and how they improve the current MDB operation. Moreover, we present with detail a specific implementation of MotivEn and the application results obtained in terms of sub-goal identification when applying it in a real robot experiment with the MDB.

ABSTRACT. In this paper we present a framework aimed at detecting emotions and sentiments in a Twitter stream. The approach uses the well-founded Latent Semantic Analysis technique, which can be seen as a bio-insipred cognitive architecture, to induce a semantic space where tweets are mapped and analysed by soft sensors. The measurements of the soft sensors are then used by a visualisation module which exploits glyphs to graphically present them. The result is an interactive map which makes easy the exploration of reactions and opinions in the whole globe regarding tweets retrieved from specific queries.

ABSTRACT. A critical level of synaptic perturbation within a trained, artificial neural system induces the nucleation of novel activation patterns, many of which could qualify as viable ideas or action plans. In building massively parallel systems requiring myriad, coupled neural modules driven in this manner, the need has arisen to selectively shift the attention of computational critics to only those portions of the neural “real estate” generating sufficiently novel activation patterns. In search of a suitable affordance to guide such attentional awareness it has been found that the rhythm of pattern generation is a quantitative indicator of the novelty of patterns being generated, and that even within a single, monolithic neural net, pattern turnover may be viewed as a superposition of frequency bands, each characterized by a temporal distribution that is a unique function of the novelty of patterns therein.

Anticipating that such synaptic perturbation is tantamount to volume neurotransmitter release within cortex, a more complete theory of cognition, creativity, and consciousness emerges.

ABSTRACT. With the fast development of internet technology, massive amounts of information have become available for people to explore. We propose an automated narrative agent that organizes and presents information to a user using storytelling techniques. It presents information by taking into consideration a combination of factors ranging from topic consistency and novelty to user interests and preferences. We present examples from using this system for generating presentations, and preliminary evaluation results.