Predictive Processing in Cognitive Robotics: A Review.

Predictive processing has become an influential framework in cognitive sciences. This framework turns the traditional view of perception upside down, claiming that the main flow of information processing is realized in a top-down, hierarchical manner. Furthermore, it aims at unifying perception, cognition, and action as a single inferential process. However, in the related literature, the predictive processing framework and its associated schemes, such as predictive coding, active inference, perceptual inference, and free-energy principle, tend to be used interchangeably. In the field of cognitive robotics, there is no clear-cut distinction on which schemes have been implemented and under which assumptions. In this letter, working definitions are set with the main aim of analyzing the state of the art in cognitive robotics research working under the predictive processing framework as well as some related nonrobotic models. The analysis suggests that, first, research in both cognitive robotics implementations and nonrobotic models needs to be extended to the study of how multiple exteroceptive modalities can be integrated into prediction error minimization schemes. Second, a relevant distinction found here is that cognitive robotics implementations tend to emphasize the learning of a generative model, while in nonrobotics models, it is almost absent. Third, despite the relevance for active inference, few cognitive robotics implementations examine the issues around control and whether it should result from the substitution of inverse models with proprioceptive predictions. Finally, limited attention has been placed on precision weighting and the tracking of prediction error dynamics. These mechanisms should help to explore more complex behaviors and tasks in cognitive robotics research under the predictive processing framework.

Live fish learn to anticipate the movement of a fish-like robot.

The ability of an individual to predict the outcome of the actions of others and to change their own behavior adaptively is called anticipation. There are many examples from mammalian species-including humans-that show anticipatory abilities in a social context, however, it is not clear to what extent fishes can anticipate the actions of their interaction partners or what the underlying mechanisms are for that anticipation. To answer these questions, we let live guppies (Poecilia reticulata) interact repeatedly with an open-loop (noninteractive) biomimetic robot that has previously been shown to be an accepted conspecific. The robot always performed the same zigzag trajectory in the experimental tank that ended in one of the corners, giving the live fish the opportunity to learn both the location of the final destination as well as the specific turning movement of the robot over three consecutive trials. The live fish's reactions were categorized into a global anticipation, which we defined as relative time to reach the robot's final corner, and a local anticipation which was the relative time and location of the live fish's turns relative to robofish turns. As a proxy for global anticipation, we found that live fish in the last trial reached the robot's destination corner significantly earlier than the robot. Overall, more than 50% of all fish arrived at the destination before the robot. This is more than a random walk model would predict and significantly more compared to all other equidistant, yet unvisited, corners. As a proxy for local anticipation, we found fish change their turning behavior in response to the robot over the course of the trials. Initially, the fish would turn after the robot, which was reversed in the end, as they began to turn slightly before the robot in the final trial. Our results indicate that live fish are able to anticipate predictably behaving social partners both in regard to final movement locations as well as movement dynamics. Given that fish have been found to exhibit consistent behavioral differences, anticipation in fish could have evolved as a mechanism to adapt to different social interaction partners.

Robots facilitate human language production.

Despite recent developments in integrating autonomous and human-like robots into many aspects of everyday life, social interactions with robots are still a challenge. Here, we focus on a central tool for social interaction: verbal communication. We assess the extent to which humans co-represent (simulate and predict) a robot's verbal actions. During a joint picture naming task, participants took turns in naming objects together with a social robot (Pepper, Softbank Robotics). Previous findings using this task with human partners revealed internal simulations on behalf of the partner down to the level of selecting words from the mental lexicon, reflected in partner-elicited inhibitory effects on subsequent naming. Here, with the robot, the partner-elicited inhibitory effects were not observed. Instead, naming was facilitated, as revealed by faster naming of word categories co-named with the robot. This facilitation suggests that robots, unlike humans, are not simulated down to the level of lexical selection. Instead, a robot's speaking appears to be simulated at the initial level of language production where the meaning of the verbal message is generated, resulting in facilitated language production due to conceptual priming. We conclude that robots facilitate core conceptualization processes when humans transform thoughts to language during speaking.

Bringing Together Robotics, Neuroscience, and Psychology: Lessons Learned From an Interdisciplinary Project.

The diversified methodology and expertise of interdisciplinary research teams provide the opportunity to overcome the limited perspectives of individual disciplines. This is particularly true at the interface of Robotics, Neuroscience, and Psychology as the three fields have quite different perspectives and approaches to offer. Nonetheless, aligning backgrounds and interdisciplinary expectations can present challenges due to varied research cultures and practices. Overcoming these challenges stands at the beginning of each productive collaboration and thus is a mandatory step in cognitive neurorobotics. In this article, we share eight lessons that we learned from our ongoing interdisciplinary project on human-robot and robot-robot interaction in social settings. These lessons provide practical advice for scientists initiating interdisciplinary research endeavors. Our advice can help to avoid early problems and deal with differences between research fields, prepare for and anticipate challenges, align project expectations, and speed up research progress, thus promoting effective interdisciplinary research across Robotics, Neuroscience, and Psychology.

Sparse coding with a somato-dendritic rule.

Cortical neurons are silent most of the time: sparse activity enables low-energy computation in the brain, and promises to do the same in neuromorphic hardware. Beyond power efficiency, sparse codes have favourable properties for associative learning, as they can store more information than local codes but are easier to read out than dense codes. Auto-encoders with a sparse constraint can learn sparse codes, and so can single-layer networks that combine recurrent inhibition with unsupervised Hebbian learning. But the latter usually require fast homeostatic plasticity, which could lead to catastrophic forgetting in embodied agents that learn continuously. Here we set out to explore whether plasticity at recurrent inhibitory synapses could take up that role instead, regulating both the population sparseness and the firing rates of individual neurons. We put the idea to the test in a network that employs compartmentalised inputs to solve the task: rate-based dendritic compartments integrate the feedforward input, while spiking integrate-and-fire somas compete through recurrent inhibition. A somato-dendritic learning rule allows somatic inhibition to modulate nonlinear Hebbian learning in the dendrites. Trained on MNIST digits and natural images, the network discovers independent components that form a sparse encoding of the input and support linear decoding. These findings confirm that intrinsic homeostatic plasticity is not strictly required for regulating sparseness: inhibitory synaptic plasticity can have the same effect. Our work illustrates the usefulness of compartmentalised inputs, and makes the case for moving beyond point neuron models in artificial spiking neural networks.

Prerequisites for an Artificial Self.

Traditionally investigated in philosophy, body ownership and agency-two main components of the minimal self-have recently gained attention from other disciplines, such as brain, cognitive and behavioral sciences, and even robotics and artificial intelligence. In robotics, intuitive human interaction in natural and dynamic environments becomes more and more important, and requires skills such as self-other distinction and an understanding of agency effects. In a previous review article, we investigated studies on mechanisms for the development of motor and cognitive skills in robots (Schillaci et al., 2016). In this review article, we argue that these mechanisms also build the foundation for an understanding of an artificial self. In particular, we look at developmental processes of the minimal self in biological systems, transfer principles of those to the development of an artificial self, and suggest metrics for agency and body ownership in an artificial self.

A Note on the Depth-from-Defocus Mechanism of Jumping Spiders.

Jumping spiders are capable of estimating the distance to their prey relying only on the information from one of their main eyes. Recently, it has been shown that jumping spiders perform this estimation based on image defocus cues. In order to gain insight into the mechanisms involved in this blur-to-distance mapping as performed by the spider and to judge whether inspirations can be drawn from spider vision for depth-from-defocus computer vision algorithms, we constructed a three-dimensional (3D) model of the anterior median eye of the Metaphidippus aeneolus, a well studied species of jumping spider. We were able to study images of the environment as the spider would see them and to measure the performances of a well known depth-from-defocus algorithm on this dataset. We found that the algorithm performs best when using images that are averaged over the considerable thickness of the spider's receptor layers, thus pointing towards a possible functional role of the receptor thickness for the spider's depth estimation capabilities.