A space of goals: the cognitive geometry of informationally bounded agents.

Traditionally, Euclidean geometry is treated by scientists as a priori and objective. However, when we take the position of an agent, the problem of selecting a best route should also factor in the abilities of the agent, its embodiment and particularly its cognitive effort. In this paper, we consider geometry in terms of travel between states within a world by incorporating information processing costs with the appropriate spatial distances. This induces a geometry that increasingly differs from the original geometry of the given world as information costs become increasingly important. We visualize this 'cognitive geometry' by projecting it onto two- and three-dimensional spaces showing distinct distortions reflecting the emergence of epistemic and information-saving strategies as well as pivot states. The analogies between traditional cost-based geometries and those induced by additional informational costs invite a generalization of the notion of geodesics as cheapest routes towards the notion of infodesics. In this perspective, the concept of infodesics is inspired by the property of geodesics that, travelling from a given start location to a given goal location along a geodesic, not only the goal, but all points along the way are visited at optimal cost from the start.

When unsupervised training benefits category learning.

Humans continuously categorise inputs, but only rarely receive explicit feedback as to whether or not they are correct. This implies that they may be integrating unsupervised information together with their sparse supervised data - a form of semi-supervised learning. However, experiments testing semi-supervised learning are rare, and are bedevilled with conflicting results about whether the unsupervised information affords any benefit. Here, we suggest that one important factor that has been paid insufficient attention is the alignment between subjects' internal representations of the stimulus material and the experimenter-defined representations that determine success in the tasks. Subjects' representations are shaped by prior biases and experience, and unsupervised learning can only be successful if the alignment suffices. Otherwise, unsupervised learning might harmfully strengthen incorrect assumptions. To test this hypothesis, we conducted an experiment in which subjects initially categorise items along a salient, but task-irrelevant, dimension, and only recover the correct categories when sufficient feedback draws their attention to the subtle, task-relevant, stimulus dimensions. By withdrawing feedback at different stages along this learning curve, we tested whether unsupervised learning improves or worsens performance when internal stimulus representations and task are sufficiently or insufficiently aligned, respectively. Our results demonstrate that unsupervised learning can indeed have opposing effects on subjects' learning. We also discuss factors limiting the degree to which such effects can be predicted from momentary performance. Our work implies that predicting and understanding human category learning in particular tasks requires assessment and consideration of the representational spaces that subjects entertain for the materials involved in those tasks. These considerations not only apply to studies in the lab, but could also help improve the design of tutoring systems and instruction.

Emerging category representation in the visual forebrain hierarchy of pigeons (Columba livia).

Recognizing and categorizing visual stimuli are cognitive functions vital for survival, and an important feature of visual systems in primates as well as in birds. Visual stimuli are processed along the ventral visual pathway. At every stage in the hierarchy, neurons respond selectively to more complex features, transforming the population representation of the stimuli. It is therefore easier to read-out category information in higher visual areas. While explicit category representations have been observed in the primate brain, less is known on equivalent processes in the avian brain. Even though their brain anatomies are radically different, it has been hypothesized that visual object representations are comparable across mammals and birds. In the present study, we investigated category representations in the pigeon visual forebrain using recordings from single cells responding to photographs of real-world objects. Using a linear classifier, we found that the population activity in the visual associative area mesopallium ventrolaterale (MVL) distinguishes between animate and inanimate objects, although this distinction is not required by the task. By contrast, a population of cells in the entopallium, a region that is lower in the hierarchy of visual areas and that is related to the primate extrastriate cortex, lacked this information. A model that pools responses of simple cells, which function as edge detectors, can account for the animate vs. inanimate categorization in the MVL, but performance in the model is based on different features than in MVL. Therefore, processing in MVL cells is very likely more abstract than simple computations on the output of edge detectors.

Forget-me-some: General versus special purpose models in a hierarchical probabilistic task.

Humans build models of their environments and act according to what they have learnt. In simple experimental environments, such model-based behaviour is often well accounted for as if subjects are ideal Bayesian observers. However, more complex probabilistic tasks require more sophisticated forms of inference that are sufficiently computationally and statistically taxing as to demand approximation. Here, we study properties of two approximation schemes in the context of a serial reaction time task in which stimuli were generated from a hierarchical Markov chain. One, pre-existing, scheme was a generically powerful variational method for hierarchical inference which has recently become popular as an account of psychological and neural data across a wide swathe of probabilistic tasks. A second, novel, scheme was more specifically tailored to the task at hand. We show that the latter model fit significantly better than the former. This suggests that our subjects were sensitive to many of the particular constraints of a complex behavioural task. Further, the tailored model provided a different perspective on the effects of cholinergic manipulations in the task. Neither model fit the behaviour on more complex contingencies that competently. These results illustrate the benefits and challenges that come with the general and special purpose modelling approaches and raise important questions of how they can advance our current understanding of learning mechanisms in the brain.

Are baboons learning "orthographic" representations? Probably not.

The ability of Baboons (papio papio) to distinguish between English words and nonwords has been modeled using a deep learning convolutional network model that simulates a ventral pathway in which lexical representations of different granularity develop. However, given that pigeons (columba livia), whose brain morphology is drastically different, can also be trained to distinguish between English words and nonwords, it appears that a less species-specific learning algorithm may be required to explain this behavior. Accordingly, we examined whether the learning model of Rescorla and Wagner, which has proved to be amazingly fruitful in understanding animal and human learning could account for these data. We show that a discrimination learning network using gradient orientation features as input units and word and nonword units as outputs succeeds in predicting baboon lexical decision behavior-including key lexical similarity effects and the ups and downs in accuracy as learning unfolds-with surprising precision. The models performance, in which words are not explicitly represented, is remarkable because it is usually assumed that lexicality decisions, including the decisions made by baboons and pigeons, are mediated by explicit lexical representations. By contrast, our results suggest that in learning to perform lexical decision tasks, baboons and pigeons do not construct a hierarchy of lexical units. Rather, they make optimal use of low-level information obtained through the massively parallel processing of gradient orientation features. Accordingly, we suggest that reading in humans first involves initially learning a high-level system building on letter representations acquired from explicit instruction in literacy, which is then integrated into a conventionalized oral communication system, and that like the latter, fluent reading involves the massively parallel processing of the low-level features encoding semantic contrasts.

Categories in the pigeon brain: A reverse engineering approach.

Pigeons are well known for their visual capabilities as well as their ability to categorize visual stimuli at both the basic and superordinate level. We adopt a reverse engineering approach to study categorization learning: Instead of training pigeons on predefined categories, we simply present stimuli and analyze neural output in search of categorical clustering on a solely neural level. We presented artificial stimuli, pictorial and grating stimuli, to pigeons without the need of any differential behavioral responding while recording from the nidopallium frontolaterale (NFL), a higher visual area in the avian brain. The pictorial stimuli differed in color and shape; the gratings differed in spatial frequency and amplitude. We computed representational dissimilarity matrices to reveal categorical clustering based on both neural data and pecking behavior. Based on neural output of the NFL, pictorial and grating stimuli were differentially represented in the brain. Pecking behavior showed a similar pattern, but to a lesser extent. A further subclustering within pictorial stimuli according to color and shape, and within gratings according to frequency and amplitude, was not present. Our study gives proof-of-concept that this reverse engineering approach-namely reading out categorical information from neural data--can be quite helpful in understanding the neural underpinnings of categorization learning.