Choice reaching with a LEGO arm robot (CoRLEGO): The motor system guides visual attention to movement-relevant information

Continuous action with a neurobiologically inspired computational approach reveals the dynamics of selection history

Everyday perception-action interaction often requires selection of a single goal from multiple possibilities. According to a recent framework of attentional control, object selection is guided not only by the well-established factors of perceptual salience and current goals but also by selection history. Yet, underlying mechanisms linking selection history and visually-guided actions are poorly understood. To examine such interplay and disentangle the impact of target and distractor history on action selection, we employed a priming-of-popout (PoP) paradigm combined with continuous tracking of reaching movements and computational modeling. Participants reached an odd-colored target among homogeneous distractors while we systematically manipulated the sequence of target and distractor colors from one trial to the next. We observed that current reach movements were significantly influenced by the interaction between attraction by the prior target feature and repulsion by the prior distractor feature. With principal component regression, we found that inhibition led by prior distractors influenced reach target selection earlier than facilitation led by the prior target. In parallel, our newly developed computational model validated that current reach target selection can be explained best by the mechanism postulating the preceded impact of previous distractors followed by a previous target. Such converging empirical and computational evidence suggests that the prior selection history triggers a dynamic interplay between target facilitation and distractor inhibition to guide goal-directed action successfully. This, in turn, highlights the necessity of an explicitly integrated approach to determine how visual attentional selection links with adaptive actions in a complex environment.

A failure to learn object shape geometry: Implications for convolutional neural networks as plausible models of biological vision

Here we examine the plausibility of deep convolutional neural networks (CNNs) as a theoretical framework for understanding biological vision in the context of image classification. Recent work on object recognition in human vision has shown that both global, and local, shape information is computed, and integrated, early during perceptual processing. Our goal was to compare the similarity in how object shape information is processed by CNNs and human observers. We tested the hypothesis that, unlike the human system, CNNs do not compute representations of global and local object geometry during image classification. To do so, we trained and tested six CNNs (AlexNet, VGG-11, VGG-16, ResNet-18, ResNet-50, GoogLeNet), and human observers, to discriminate geometrically possible and impossible objects. The ability to complete this task requires computation of a representational structure of shape that encodes both global and local object geometry because the detection of impossibility derives from an incongruity between well-formed local feature conjunctions and their integration into a geometrically well-formed 3D global shape. Unlike human observers, none of the tested CNNs could reliably discriminate between possible and impossible objects. Detailed analyses using gradient-weighted class activation mapping (GradCam) of CNN image feature processing showed that network classification performance was not constrained by object geometry. In contrast, if classification could be made based solely on local feature information in line drawings the CNNs were highly accurate. We argue that these findings reflect fundamental differences between CNNs and human vision in terms of underlying image processing structure. Notably, unlike human vision, CNNs do not compute representations of object geometry. The results challenge the plausibility of CNNs as a framework for understanding image classification in biological vision systems.

Autonomous Sequence Generation for a Neural Dynamic Robot: Scene Perception, Serial Order, and Object-Oriented Movement

Neurally inspired robotics already has a long history that includes reactive systems emulating reflexes, neural oscillators to generate movement patterns, and neural networks as trainable filters for high-dimensional sensory information. Neural inspiration has been less successful at the level of cognition. Decision-making, planning, building and using memories, for instance, are more often addressed in terms of computational algorithms than through neural process models. To move neural process models beyond reactive behavior toward cognition, the capacity to autonomously generate sequences of processing steps is critical. We review a potential solution to this problem that is based on strongly recurrent neural networks described as neural dynamic systems. Their stable states perform elementary motor or cognitive functions while coupled to sensory inputs. The state of the neural dynamics transitions to a new motor or cognitive function when a previously stable neural state becomes unstable. Only when a neural robotic system is capable of acting autonomously does it become a useful to a human user. We demonstrate how a neural dynamic architecture that supports autonomous sequence generation can engage in such interaction. A human user presents colored objects to the robot in a particular order, thus defining a serial order of color concepts. The user then exposes the system to a visual scene that contains the colored objects in a new spatial arrangement. The robot autonomously builds a scene representation by sequentially bringing objects into the attentional foreground. Scene memory updates if the scene changes. The robot performs visual search and then reaches for the objects in the instructed serial order. In doing so, the robot generalizes across time and space, is capable of waiting when an element is missing, and updates its action plans online when the scene changes. The entire flow of behavior emerges from a time-continuous neural dynamics without any controlling or supervisory algorithm.

Excitatory versus inhibitory feedback in Bayesian formulations of scene construction

The selective attention for identification model (SAIM) is an established model of selective visual attention. SAIM implements translation-invariant object recognition, in scenes with multiple objects, using the parallel distributed processing (PDP) paradigm. Here, we show that SAIM can be formulated as Bayesian inference. Crucially, SAIM uses excitatory feedback to combine top-down information (i.e. object knowledge) with bottom-up sensory information. By contrast, predictive coding implementations of Bayesian inference use inhibitory feedback. By formulating SAIM as a predictive coding scheme, we created a new version of SAIM that uses inhibitory feedback. Simulation studies showed that both types of architectures can reproduce the response time costs induced by multiple objects-as found in visual search experiments. However, due to the different nature of the feedback, the two SAIM schemes make distinct predictions about the motifs of microcircuits mediating the effects of top-down afferents. We discuss empirical (neuroimaging) methods to test the predictions of the two inference architectures.

Task instructions modulate unit-decade binding in two-digit number representation

Previous studies have found decomposed processes, as well as holistic processes, in the representation of two-digit numbers. The present study investigated the influence of task instruction on such processes. Participants completed both magnitude and parity tasks in one of three instructional conditions, where they were asked to either consider two-digit numbers as a whole or to focus on one specific digit. In two experiments, we found that when participants were asked to consider the two digits as an integrated number, they always exhibited a unit-decade compatibility effect, indicating a failure of selective attention on the digit relevant to the given task. However, the mere presence of the neighboring digit is not a sufficient condition for the compatibility effect: when participants were explicitly asked to process a specific digit, their success/failure to selectively ignore the irrelevant digit depended on task requirements. Further, computer mouse tracking indicated that the locus of the compatibility effect was related to late response-related processing. The results signify the deep involvement of top-down processes in unit-decade binding for two-digit number representation.

Abandoning and modifying one action plan for alternatives

Visual scenes are often complex and crowded with many different objects. To interact effectively, we must choose one object at a time as a goal for action. Certain external cues can act as a stop signal, quickly cancelling an ongoing action. Less recognized are internal signals. These can come from recent experience, anticipated action outcomes, cognitive states, and when attention is captured by a salient object. These signals elevate one action plan over alternatives and can quickly modify an initial choice. Here, we focus on these internal processes responsible for selecting, abandoning and modifying action plans. We first highlight how the brain resolves competition among multiple action plans. Critical is the existence of parallel motor planning processes, which allow efficient and timely changes. Then, we discuss how the action system interplays with perception, attention and memory processes to bias action selection and suppress or modify erroneous selections. Subsequently, we show how tracking the continuous modification of action trajectories can provide a tool to read out changes in internal cognitive states. Taken together, we shed light on a broader view that sensorimotor networks can continuously modify actions through simultaneous evaluation of alternative activities in concert with widely distributed perceptual and cognitive networks.This article is part of the themed issue 'Movement suppression: brain mechanisms for stopping and stillness'.

Reversing the Manual Digit Bias in Two-Digit Number Comparison

Though recent work in numerical cognition has supported a strong tie between numerical and spatial representations (e.g., a mental number line), less is known about such ties in multi-digit number representations. Along this line, Bloechle, Huber, and Moeller (2015) found that pointing positions in two-digit number comparison were biased leftward toward the decade digit. Moreover, this bias was reduced in unit-decade incompatible pairs. In the present study, we tracked computer mouse movements as participants compared two-digit numbers to a fixed standard (55). Similar to Bloechle et al. (2015) , we found that trajectories exhibited a leftward bias that was reduced for unit-decade incompatible comparisons. However, when positions of response labels were reversed, the biases reversed. That is, we found a rightward bias for compatible pairs that was reduced for incompatible pairs. This result calls into question a purely embodied representation of place value structure and instead supports a competition model of two-digit number representation.

A Neural Dynamic Architecture for Reaching and Grasping Integrates Perception and Movement Generation and Enables On-Line Updating

Reaching for objects and grasping them is a fundamental skill for any autonomous robot that interacts with its environment. Although this skill seems trivial to adults, who effortlessly pick up even objects they have never seen before, it is hard for other animals, for human infants, and for most autonomous robots. Any time during movement preparation and execution, human reaching movement are updated if the visual scene changes (with a delay of about 100 ms). The capability for online updating highlights how tightly perception, movement planning, and movement generation are integrated in humans. Here, we report on an effort to reproduce this tight integration in a neural dynamic process model of reaching and grasping that covers the complete path from visual perception to movement generation within a unified modeling framework, Dynamic Field Theory. All requisite processes are realized as time-continuous dynamical systems that model the evolution in time of neural population activation. Population level neural processes bring about the attentional selection of objects, the estimation of object shape and pose, and the mapping of pose parameters to suitable movement parameters. Once a target object has been selected, its pose parameters couple into the neural dynamics of movement generation so that changes of pose are propagated through the architecture to update the performed movement online. Implementing the neural architecture on an anthropomorphic robot arm equipped with a Kinect sensor, we evaluate the model by grasping wooden objects. Their size, shape, and pose are estimated from a neural model of scene perception that is based on feature fields. The sequential organization of a reach and grasp act emerges from a sequence of dynamic instabilities within a neural dynamics of behavioral organization, that effectively switches the neural controllers from one phase of the action to the next. Trajectory formation itself is driven by a dynamical systems version of the potential field approach. We highlight the emergent capacity for online updating by showing that a shift or rotation of the object during the reaching phase leads to the online adaptation of the movement plan and successful completion of the grasp.

The Emergence of Selective Attention through Probabilistic Associations between Stimuli and Actions

In this paper we show how a multilayer neural network trained to master a context-dependent task in which the action co-varies with a certain stimulus in a first context and with a second stimulus in an alternative context exhibits selective attention, i.e. filtering out of irrelevant information. This effect is rather robust and it is observed in several variations of the experiment in which the characteristics of the network as well as of the training procedure have been varied. Our result demonstrates how the filtering out of irrelevant information can originate spontaneously as a consequence of the regularities present in context-dependent training set and therefore does not necessarily depend on specific architectural constraints. The post-evaluation of the network in an instructed-delay experimental scenario shows how the behaviour of the network is consistent with the data collected in neuropsychological studies. The analysis of the network at the end of the training process indicates how selective attention originates as a result of the effects caused by relevant and irrelevant stimuli mediated by context-dependent and context-independent bidirectional associations between stimuli and actions that are extracted by the network during the learning.

Developing Dynamic Field Theory Architectures for Embodied Cognitive Systems with cedar

Embodied artificial cognitive systems, such as autonomous robots or intelligent observers, connect cognitive processes to sensory and effector systems in real time. Prime candidates for such embodied intelligence are neurally inspired architectures. While components such as forward neural networks are well established, designing pervasively autonomous neural architectures remains a challenge. This includes the problem of tuning the parameters of such architectures so that they deliver specified functionality under variable environmental conditions and retain these functions as the architectures are expanded. The scaling and autonomy problems are solved, in part, by dynamic field theory (DFT), a theoretical framework for the neural grounding of sensorimotor and cognitive processes. In this paper, we address how to efficiently build DFT architectures that control embodied agents and how to tune their parameters so that the desired cognitive functions emerge while such agents are situated in real environments. In DFT architectures, dynamic neural fields or nodes are assigned dynamic regimes, that is, attractor states and their instabilities, from which cognitive function emerges. Tuning thus amounts to determining values of the dynamic parameters for which the components of a DFT architecture are in the specified dynamic regime under the appropriate environmental conditions. The process of tuning is facilitated by the software framework cedar, which provides a graphical interface to build and execute DFT architectures. It enables to change dynamic parameters online and visualize the activation states of any component while the agent is receiving sensory inputs in real time. Using a simple example, we take the reader through the workflow of conceiving of DFT architectures, implementing them on embodied agents, tuning their parameters, and assessing performance while the system is coupled to real sensory inputs.