Will Penny

Factorial designs are a mainstay of the scientific paradigm, allowing the effects of multiple experimental factors and their interactions to be efficiently studied within a single experiment. In brain imaging, however, multivariate data analyses commonly proceed using multivariate decoding and we argue that the standard “difference of accuracies” test is not a true test of interactions. To remedy this situation we propose an encoding method based on a Bayesian multivariate linear model which is ideally suited to the factorial analysis of such data. We show how it can be used to test for multivariate main effects and interactions using data from EEG studies of reward learning and declarative memory. This approach additionally allows for null hypotheses to be accepted and allows one to infer whether multivariate effects are driven by collections of univariate effects or voxel dependencies. We also propose that those wishing to pursue tests for interactions using decoding methods use an “accuracy of differences” test.

Factorial designs are a mainstay of the scientific paradigm, allowing the effects of multiple experimental factors and their interactions to be efficiently studied within a single experiment. In brain imaging, however, multivariate data analyses commonly proceed using multivariate decoding and we argue that the standard “difference of accuracies” test is not a true test of interactions. To remedy this situation we propose an encoding method based on a Bayesian Multivariate Linear model which is ideally suited to the factorial analysis of such data. We show how it can be used to test for multivariate main effects and interactions using data from EEG studies of Reward Learning and Declarative Memory. This approach additionally allows for null hypotheses to be accepted and allows one to infer whether multivariate effects are driven by collections of univariate effects or voxel dependencies. We also propose that those wishing to pursue tests for interactions using decoding methods use an “accuracy of differences” test. Author summary: Our understanding of human brain function has been transformed by non-invasive imaging methods such as functional Magnetic Resonance Imaging and Electroencephalography. Statistical modelling has been central to this endeavour with foundational work employing a univariate encoding method in which multiple characteristics of participants behaviour (the stimuli they are exposed to, the decisions they make, the events they remember) are used to predict brain activity at a single point or voxel element (pixel/voxel) in a brain recording, and this process repeats for all voxels in an image. Subsequent work has used multivariate decoding methods which identify categorical behavioural variables from multivariate brain imaging data. Here we propose a new type of multivariate encoding approach in which a Bayesian multivariate linear model is used to predict multivariate image data from multivariate behavioural variables. This approach has three advantages (i) it can correctly test for interactions among experimental factors, (ii) we can quantify the experimental evidence in support of no experimental effects being present and (iii) we can make inferences about the nature of the multivariate effects.

This paper investigates models of working memory in which memory traces evolve according to stochastic attractor dynamics. These models have previously been shown to account for response-biases that are manifest across multiple trials of a visual working memory task. Here we adapt this approach by making the stable fixed points correspond to the multiple items to be remembered within a single-trial, in accordance with standard dynamical perspectives of memory, and find evidence that this multi-item model can provide a better account of behavioural data from continuous-report tasks. Additionally, the multi-item model proposes a simple mechanism by which swap-errors arise: memory traces diffuse away from their initial state and are captured by the attractors of other items. Swap-error curves reveal the evolution of this process as a continuous function of time throughout the maintenance interval and can be inferred from experimental data. Consistent with previous findings, we find that empirical memory performance is not well characterised by a purely-diffusive process but rather by a stochastic process that also embodies error-correcting dynamics.

This paper proposes a model of hippocampal-based category learning using a cluster-based inference framework which produces two systems (i) an extendable cluster-based memory module (CMM) that provides representations of learning episodes with strong pattern separation, and supports online decision making during that learning, (ii) cluster-based task modules (CTMs) which consolidate clusters learnt by CMM to support higher-accuracy decision making in the mid-term. Learning in CMMs optimises the joint probability of stimuli and category labels, whereas learning in CTMs optimises the probability of categories given stimuli. The latter system learns from the former via a process of “cluster consolidation”. We apply the model to data from a behavioral learning task and find that, as well as improving categorisation performance, cluster consolidation decreases recognition scores for old cues but increases them for novel cues. This model-based perspective explains forgetting and false memory effects as serving future categorisation performance. The paper also expresses a view that models of memorybased cognition should provide human-level performance on complex categorisation tasks, and do so with minimal labelled data. In working toward this goal we therefore compared the accuracy of CMM- and CTM-based decision making to standard Softmax Regression approaches on benchmark machine learning datasets. This produced mixed results. We found some significant improvements of CMM over Softmax Regression and of CTM over CMM. Overall, our framework reconciles complementary learning systems theory with more recent findings from cognitive neuroscience of generative replay and hippocampal-based generalisation.

This paper uses constructs from machine learning to define pairs of learning tasks that either shared or did not share a common subspace. Human subjects then learnt these tasks using a feedback-based approach and we hypothesised that learning would be boosted for shared subspaces. Our findings broadly supported this hypothesis with either better performance on the second task if it shared the same subspace as the first, or positive correlations over task performance for shared subspaces. These empirical findings were compared to the behaviour of a Neural Network model trained using sequential Bayesian learning and human performance was found to be consistent with a minimal capacity variant of this model. Networks with an increased representational capacity, and networks without Bayesian learning, did not show these transfer effects. We propose that the concept of shared subspaces provides a useful framework for the experimental study of human multitask and transfer learning.

There is consensus that activation within distributed functional brain networks underlies human thought. The impact of this consensus is limited, however, by a gap that exists between data-driven correlational analyses that specify where functional brain activity is localized using functional magnetic resonance imaging (fMRI), and neural process accounts that specify how neural activity unfolds through time to give rise to behavior. Here, we show how an integrative cognitive neuroscience approach may bridge this gap. In an exemplary study of visual working memory, we use multilevel Bayesian statistics to demonstrate that a neural dynamic model simultaneously explains behavioral data and predicts localized patterns of brain activity, outperforming standard analytic approaches to fMRI. The model explains performance on both correct trials and incorrect trials where errors in change detection emerge from neural fluctuations amplified by neural interaction. Critically, predictions of the model run counter to cognitive theories of the origin of errors in change detection. Results reveal neural patterns predicted by the model within regions of the dorsal attention network that have been the focus of much debate. The model-based analysis suggests that key areas in the dorsal attention network such as the intraparietal sulcus play a central role in change detection rather than working memory maintenance, counter to previous interpretations of fMRI studies. More generally, the integrative cognitive neuroscience approach used here establishes a framework for directly testing theories of cognitive and brain function using the combined power of behavioral and fMRI data. (PsycInfo Database Record (c) 2021 APA, all rights reserved).

This paper uses constructs from the field of multitask machine learning to define pairs of learning tasks that either shared or did not share a common subspace. Human subjects then learnt these tasks using a feedback-based approach. We found, as hypothesised, that subject performance was significantly higher on the second task if it shared the same subspace as the first, an advantage that played out most strongly at the beginning of the second task. Additionally, accuracy was positively correlated over subjects learning same-subspace tasks but was not correlated for those learning different-subspace tasks. These results, and other aspects of learning dynamics, were compared to the behaviour of a Neural Network model trained using sequential Bayesian inference. Human performance was found to be consistent with a Soft Parameter Sharing variant of this model that constrained representations to be similar among tasks but only when this aided learning. We propose that the concept of shared subspaces provides a useful framework for the experimental study of human multitask and transfer learning. Author summary How does knowledge gained from previous experience affect learning of new tasks ? This question of “Transfer Learning” has been addressed by teachers, psychologists, and more recently by researchers in the fields of neural networks and machine learning. Leveraging constructs from machine learning, we designed pairs of learning tasks that either shared or did not share a common subspace. We compared the dynamics of transfer learning in humans with those of a multitask neural network model, finding that human performance was consistent with a soft parameter sharing variant of the model. Learning was boosted in the early stages of the second task if the same subspace was shared between tasks. Additionally, accuracy between tasks was positively correlated but only when they shared the same subspace. Our results highlight the roles of subspaces, showing how they could act as a learning boost if shared, and be detrimental if not.

In generative modeling of neuroimaging data, such as dynamic causal modeling (DCM), one typically considers several alternative models, either to determine the most plausible explanation for observed data (Bayesian model selection) or to account for model uncertainty (Bayesian model averaging). Both procedures rest on estimates of the model evidence, a principled trade-off between model accuracy and complexity. In DCM, the log evidence is usually approximated using variational Bayes (VB) under the Laplace approximation (VBL). Although this approach is highly efficient, it makes distributional assumptions and can be vulnerable to local extrema. An alternative to VBL is Markov Chain Monte Carlo (MCMC) sampling, which is asymptotically exact but orders of magnitude slower than VB. This has so far prevented its routine use for DCM. This paper makes four contributions. First, we introduce a powerful MCMC scheme – thermodynamic integration (TI) – to neuroimaging and present a derivation that establishes a theoretical link to VB. Second, this derivation is based on a tutorial-like introduction to concepts of free energy in physics and statistics. Third, we present an implementation of TI for DCM that rests on population MCMC. Fourth, using simulations and empirical functional magnetic resonance imaging (fMRI) data, we compare log evidence estimates obtained by TI, VBL, and other MCMC-based estimators (prior arithmetic mean and posterior harmonic mean). We find that model comparison based on VBL gives reliable results in most cases, justifying its use in standard DCM for fMRI. Furthermore, we demonstrate that for complex and/or nonlinear models, TI may provide more robust estimates of the log evidence. Importantly, accurate estimates of the model evidence can be obtained with TI in acceptable computation time. This paves the way for using DCM in scenarios where the robustness of single-subject inference and model selection becomes paramount, such as differential diagnosis in clinical applications.

Dynamic causal modeling (DCM) is a framework for making inferences about changes in brain connectivity using neuroimaging data. We fitted DCMs to high-density EEG data from subjects performing a semantic picture matching task. The subjects are carrie