On the Rational Bounds of Cognitive Control

Artificial systems currently outperform humans in diverse computational domains, but none has achieved parity in speed and overall versatility of mastering novel tasks. A critical component to human success, in this regard, is the ability to redeploy and redirect data passed between cognitive subsystems (via abstract feature representations) in response to changing task demands. However, analyzing shared representations is difficult in neural systems with distributed nonlinear coding. This work presents a simple but effective approach to this problem. In experiments, the proposed model robustly predicts behavior and performance of multitasking networks on natural language data (MNIST) using common deep network architectures. Consistent with existing theory in cognitive control, representation structure varies in response to (a) environmental pressures for representation sharing, (b) demands for parallel processing capacity, and (c) tolerance for crosstalk. Implications for geometric (dimension, curvature), functional (automaticity, generalizability, modularity), and applied aspects of representation learning are discussed.

Regression, encoding, control: an integrated approach to shared representations with distributed coding

The integration of behavioral phenomena into mechanistic models of cognitive function is a fundamental staple of cognitive science. Yet, researchers are beginning to accumulate increasing amounts of data without having the temporal or monetary resources to integrate these data into scientific theories. We seek to overcome these limitations by incorporating existing machine learning techniques into an open-source pipeline for the automated construction of quantitative models. This pipeline leverages the use of neural architecture search to automate the discovery of interpretable model architectures, and automatic differentiation to automate the fitting of model parameters to data. We evaluate the utility of these methods based on their ability to recover quantitative models of human information processing from synthetic data. We find that these methods are capable of recovering basic quantitative motifs from models of psychophysics, learning and decision making. We also highlight weaknesses of this framework and discuss future directions for their mitigation.

Recovering Quantitative Models of Human Information Processing with Differentiable Architecture Search

Current models of cognitive control address selection among tasks in terms of a cost-benefit tradeoff. Importantly, they usually assume a fixed level of competence for each candidate task when estimating its value. However, performing a task can improve competence through learning, which should be factored into estimates of future value. Here, we consider an extension of the Expected Value of Control (EVC) theory that includes such estimates. We demonstrate that control allocation is a function of task learnability, and show the use of this model by generating novel predictions in cognitive effort discounting tasks. We argue that the value of learning in control allocation may account for the seemingly paradoxical finding that sometimes participants prefer more difficult (i.e., costly) tasks, and discuss how the model can be leveraged to further our understanding of human decision making and cognitive impairments.

The Value of Learning and Cognitive Control Allocation

There is major potential for using electroencephalography (EEG) in brain decoding that has been untapped due to the need for large amounts of data. Advances in machine learning have mitigated this need through data augmentation techniques, such as Generative Adversarial Networks (GANs). Here, we gauged the extent to which GANs can augment EEG data to enhance classification performance. Our objectives were to determine which classifiers benefit from GAN-augmented EEG and to estimate the impact of sample sizes on GAN-enhancements. We investigated three classifiers---neural networks, support vector machines, and logistic regressions---across seven sample sizes ranging from 5 to 100 participants. GAN-augmented EEG enhanced classification for neural networks and support vector machines, but not logistic regressions. Further, GAN-enhancements diminished as sample sizes increased---suggesting it is most effective with small samples, which may facilitate research that is unable to collect large amounts of data, such as with clinical populations.

Augmenting EEG with Generative Adversarial Networks Enhances Brain Decoding Across Classifiers and Sample Sizes

One possibility that may facilitate learning difficult tasks is building a specific curriculum, i.e., to start with simpler tasks and continue with increasing task difficulty. While such curriculum learning has long been considered and investigated in educational settings, little research has been conducted on curriculum learning in well-controlled experimental settings. To fill this gap, we investigated curriculum learning in a perceptual decision-making task. Across two experiments, we tested whether different training conditions (ascending difficulty, descending difficulty, random difficulty, easy difficulty, and difficult difficulty) affected participants' performance in a subsequent testing phase. We found that participants trained on a sequence of varying difficulty (i.e., ascending difficulty, descending difficulty, and random difficulty) outperformed others from easy-only and difficult-only training. However, we found no differences during testing between the conditions with varying difficulty. These results suggest that varying the difficulty during training yields advantages when learning the random-dot-motion task.

Sebastian Musslick

6

26

SHORT BIO

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES