Learning and adaptation: neural and behavioural mechanisms behind behaviour change

ABSTRACT

This special issue presents perspectives on learning and adaptation as they apply to a number of cognitive phenomena including pupil dilation in humans and attention in robots, natural language acquisition and production in embodied agents (robots), human-robot game play and social interaction, neural-dynamic modelling of active perception and neural-dynamic modelling of infant development in the Piagetian A-not-B task. The aim of the special issue, through its contributions, is to highlight some of the critical neural-dynamic and behavioural aspects of learning as it grounds adaptive responses in robotic- and neural-dynamic systems.

KEYWORDS:

• Embodiment
• neural dynamics
• conditioning
• interaction
• multimodal learning
• self-organisation

1. Introduction to the special issue

Research into animal and human behaviour has provided an important source to understanding common neural mechanisms that underlie both learning and adaptation. Learning can be broadly described as producing changes within an organism enabling more effective behaviour within its environment; adaptation entails behavioural adjustments to environmental change that may or may not be the direct products of learning. In this special issue, several experts in the areas of cognitive robotics and computational modelling have contributed articles that shed light on the embodied and interactive dynamics that may bring to bear on learning and adaptation.

The contributions to this special issue concern neural modelling and learning aspects of eye movement (Johansson and Balkenius, Quinton and Goffart), neural-dynamic modelling (specifically neural field modelling, cf. Amari, 1977) of sensorimotor activity underlying cognitive phenomena (again Quinton and Goffart, Dineva and Schöner), the interactive dynamics of human-robot interactions where robots deploy “social strategies” with discussion on neural-dynamic (neural field) implementations (Barakova et al.), and finally a study of neural dynamics and behaviour in the context of language production and decomposition using a multimodal embodied (robotics) approach (Heinrich and Wermter).

Johansson and Balkenius produced a computational model of pupil dilation incorporating a range of neurophysiological components purported to be influential. An important aspect of modelling – the shape of the light reflex – was qualitatively influenced by the combination of these neurophysiological components without a requirement for strong parameter fitting. Furthermore, pupil dilation is considered, by the authors, and computationally demonstrated, to be influenced by conditioning – pictures of identifiable light sources induce pupil contraction, not just the actual light source – emotional state and cognitive processing. Finally, the authors suggest that the model of pupillary activation provides something of a template for studying other aspects of peripheral nervous system activity, an aspect of adaptive behaviour that is understudied in cognitive science.

Quinton and Goffart extend a neural field approach to the modelling of visual perception, emphasising its dynamics, sensorimotor character, and anticipatory nature. In the spirit of the active vision approach of O’Regan and Noe (2001) that considers the sensorimotor dynamics in perception, the authors introduce the notion of the Active Neural Field. This version of neural field equation provides a means for neural dynamics to not only anticipate position of a moving target in a visual tracking task, as in the authors’ previous work, but also to drive overt attention shift, initiating eye movements that aim to keep the target in the centre of the field of view. The actual movement simplifies the problem of learning to predict the target’s movement, decreasing the space that learning needs to cover. The learned predictive component, in its turn, facilitates target detection and tracking. This approach makes visual perception a truly sensorimotor process, in which, mathematically, a point attractor in a joined sensorimotor space replaces peak movement in sensory-driven neural field. The presented proof-of-concept model can account for emergence of qualitatively distinct goal-directed eye-movements driven by an interplay of sensory inputs, neuronal dynamics, movement and learning.

Many learning paradigms have emphasised the importance of the history of decision-making to adaptive behaviour (Lowe, Almér, Billiing, Sandamirskaya, & Balkenius, 2017; Pipkin & Vollmer, 2009). In this special issue, Dineva and Schöner, following the tradition of neural field modelling of the sensorimotor origins of decision-making in the A-not-B task (Spencer & Schöner, 2003; Thelen, Schöner, Scheier, & Smith, 2001), demonstrate with a neural-dynamic model the importance of considering the motor outcome of a decision when forming a memory trace. In contrast to earlier neural field models of memory trace formation in the A-not-B task (e.g. Thelen et al., 2001), the authors point out that to account for influence of spontaneous errors on subsequent decisions, it is not enough to model memory trace formation based on the sensory evidence that led to the decision. To the contrary, the motor outcome of the decision – the actual movement generated to either the cued or the erroneous location – is what seems to bias future decisions. The authors note that most models of decision making “stop” at the moment of the decision and do not consider the necessity to stabilise the decision outcome to enable actual motor response. They demonstrate the importance of including the motor output when modelling decision-making by pooling data from previous studies of the A-not-B task and analysing the statistics of spontaneous errors.

In Barakova et al. human-robot interactive dynamics were studied in a checkers game according to the level of engagement that infants perceived with a (NAO) robot inter-actor or when playing against a computer (zero-sum game). The findings were that robots enhanced the level of engagement that the infants perceived (based on questionnaire feedback) and that a social strategy employed by the robot (focusing on engagement via occasional mistake making) facilitated this level of engagement. The authors also propose that neural field models would provide suitable substitutes for the randomly set mistake-making strategy that the robots deployed in order to foster human engagement. They point out the use of neural field models in relation to infant development and learning. In the context of the Dineva and Schöner article in this issue, the “spontaneous error” and “perseverative error” that young infants make during the A-not-B task, might be incorporated into the modelling of either the robot strategy or otherwise into the human-robot interaction per se.

In Heinrich and Wermter’s contribution, a cortical-inspired model was used on a robot for embodied grounding of language acquisition. The purpose of the work was to understand better the neural and behavioural mechanisms of natural language. The authors used a continuous time (multimodal) recurrent neural network (CTRNN) for the hierarchical modelling of speech production aspects. The model used standard learning methods including Backpropagation through time and Kullback–Leibler Divergence but the adaptive nature of the robot was a function of its multimodal processing. Multimodality entailed the use of such CTRNNs to implement auditory, somatosensory and visual aspects of speech production in the robotic embodied context. Such multimodality enabled the model to decompose a contextual sequence into its primitive constituents (based on previous work, cf. Heinrich & Wermter, 2014; Heinrich, Magg, & Wermter, 2015) but also construct an abstract context through linking sequences of such primitives (phonemes).

2. Conclusion

The contributions to this special issue treat learning as an integrative process driven by both the sensorimotor dynamics coupled to the physical world and intrinsic neural dynamics. Learning emerges in (modelled or physical) interaction with a physical or social environment, but is also structurally constrained by the modelled neuronal system. The special issue aims, therefore, to furnish readers with new insights into neural processes and behavioural modes of learning and adaptation. The contributions to the issue also highlight learning and adaptation as grounding cognitive phenomena such as decision making, movement generation, and communication in interaction according to sensorimotor activity including forms thereof that are interwoven with social interaction.

Disclosure statement

No potential conflict of interest was reported by the authors.