Executive attention, task selection and attention-based learning in a neurally controlled simulated robot

Abstract

We describe the design and implementation of an integrated neural architecture, modelled on human executive attention, which is used to control both automatic (reactive) and willed action selection in a simulated robot. The model, based upon Norman and Shallice's supervisory attention system, incorporates important features of human attentional control: selection of an intended task over a more salient automatic task; priming of future tasks that are anticipated; and appropriate levels of persistence of focus of attention. Recognising that attention-based learning, mediated by the limbic system, and the hippocampus in particular, plays an important role in adaptive learning, we extend the Norman and Shallice model, introducing an intrinsic, attention-based learning mechanism that enhances the automaticity of willed actions and reduces future need for attentional effort. These enhanced features support a new level of attentional autonomy in the operation of the simulated robot. Some properties of the model are explored using lesion studies, leading to the identification of a correspondence between the behavioural pathologies of the simulated robot and those seen in human patients suffering dysfunction of executive attention. We discuss briefly the question of how executive attention may have arisen due to selective pressure. LaBerge identifies three characteristics traditionally associated with executive attention and considers that they should be evident in any model [36]. Grounding these three properties in the domain of action selection, we have: Keywords1. selection of a willed action over a more salient, automatically selected, action. Here, the notion of salience is intimately connected to environmentally derived stimuli in the degree to which they accord with the relevance of contending actions. However, it may also derive from internal/innate drives. For example, the salience of feeding behaviour is determined both by the availability of food in the environment and by a sense of hunger/satiation. Willed action selection involves the application of an internally derived attentional signal which results in the (more likely) performance of a less salient act in preference to a more salient act. The attentional effort needed to will one familiar action in place of another is usually intermittent, or even momentary. The willing of wholly unfamiliar actions may require more persistent attention.2. priming of an anticipated future action. Priming, too, is associated with an internally derived attentional signal. On this occasion, the potentiation does not result in the immediate expression of the behaviour, rather it enhances the salience of the behaviour so that, when the appropriate anticipated circumstances arise, there is a greater likelihood that the anticipated task will be selected. Priming is associated with enhanced speed of task switching.3. use of memory for sustained task focus. Memory is particularly important when resumption of a suspended task requires recall of some past state or stimulus...

A Robotic Cognitive Control Framework for Collaborative Task Execution and Learning

“…Cognitive control mechanisms have also been proposed within the the Intelligent Soft Arm Control (ISAC) architecture (Kawamura et al, 2008), where attention is mainly deployed to assign priority values to orient the focus of perception. More related to our approach, a neural SAS-based executive system for robot control in a simulated environment is proposed by Garforth et al (2006), where only simple foraging tasks are considered as a proof of concept. In contrast, we are interested in a practical framework that can scale the complexity of real-world robotic collaborative tasks in everyday scenarios.…”

“…As far as human–robot collaboration is concerned, attention models are usually deployed for visual perception and exploited for implicit nonverbal communication (Breazeal, Kidd, Thomaz, Hoffman, & Berlin, 2005; Muller & Knoll, 2009), joint attention (Scassellati, 1999), anticipation (Hoffman & Breazeal, 2007), perspective taking (Trafton et al., 2005), active perception (Breazeal et al., 2001; Demiris & Khadhouri, 2006), etc. In contrast, we propose attention mechanisms for executive control, which are rarely considered in the robot literature (Broquère et al., 2014; Garforth et al., 2006; Kawamura et al., 2008) and usually not exploited for the orchestration of concurrent structured tasks in real‐world robotics systems.…”

Controlled and Automatic Processing in Animals and Machines with Application to Autonomous Vehicle Control

“…The NS model has recently been reinvigorated in quantitative form by Garforth et al [16]. Here, key components of the NS scheme have been incorporated into a simulated autonomous robot model that 'forages for food'.…”

Events, Neural Systems and Time Series

“…In terms of neural systems, one needs to start by considering what the events might be 3 . Events could range from arrival of an action potential at a synapse (one might start at a lower level, for example with the release of a vesicle from the presynaptic terminal, or the arrival of molecules of neurotransmitter from such a vesicle at a single post-synaptic ion channel), or with the detection of some percept in the auditory or visual cortex, or the execution of a motor command.…”

“…We clearly do have (first-person) mental events, and we believe that these are mediated by action potential/synaptic events: but the relationship between these two is not yet anything like fully elucidated. There are higher-level theories of mental function (such as those of [5] or [3]), but these are largely narrative models 5 rather than precise simulations. For natural environmental systems, the issues are perhaps easier, since we already have a large volume of physics, much of which is about the inter-relationship between events in the physical world.…”