With practice, humans tend to improve their performance on most tasks. But do such improvements then generalize to new tasks? Although early work documented primarily task-specific learning outcomes in the domain of perceptual learning [1-3], an emerging body of research has shown that significant learning generalization is possible under some training conditions [4-9]. Interestingly, however, research in this vein has focused nearly exclusively on just one possible manifestation of learning generalization, wherein training on one task produces an immediate boost to performance on the new task. For instance, it is this form of generalization that is most frequently referred to when discussing learning "transfer" [10, 11]. Essentially no work in this domain has focused on a second possible manifestation of generalization, wherein the knowledge or skills acquired via training, despite not being directly applicable to the new task, nonetheless allow the new task to be learned more efficiently [12-15]. Here, in both the visual category learning and visual perceptual learning domains, we demonstrate that sequentially training participants on tasks that share a common high-level task structure can produce faster learning of new tasks, even in cases where there is no immediate benefit to performance on the new tasks. We further show that methods commonly employed in the field may fail to detect or else conflate generalization that manifests as increased learning rate with generalization that manifests as immediate boosts to performance. These results thus lay the foundation for the various routes to learning generalization to be more thoroughly explored.

No abstract

How does the human brain encode semantic information about objects? This paper reconciles two seemingly contradictory views. The first proposes that local neural populations independently encode semantic features; the second, that semantic representations arise as a dynamic distributed code that changes radically with stimulus processing. Combining simulations with a well-known neural network model of semantic memory, multivariate pattern classification, and human electrocorticography, we find that both views are partially correct: information about the animacy of a depicted stimulus is distributed across ventral temporal cortex in a dynamic code possessing feature-like elements posteriorly but with elements that change rapidly and nonlinearly in anterior regions. This pattern is consistent with the view that anterior temporal lobes serve as a deep cross-modal ‘hub’ in an interactive semantic network, and more generally suggests that tertiary association cortices may adopt dynamic distributed codes difficult to detect with common brain imaging methods.

The increasing of public neuroimaging datasets opens a door to analyzing homogeneous human brain conditions across datasets by transfer learning (TL). However, neuroimaging data are high-dimensional, noisy, and with small sample sizes. It is challenging to learn a robust model for data across different cognitive experiments and subjects. A recent TL approach minimizes domain dependence to learn common cross-domain features, via the Hilbert-Schmidt Independence Criterion (HSIC). Inspired by this approach and the multi-source TL theory, we propose a Side Information Dependence Regularization (SIDeR) learning framework for TL in brain condition decoding. Specifically, SIDeR simultaneously minimizes the empirical risk and the statistical dependence on the domain side information, to reduce the theoretical generalization error bound. We construct 17 brain decoding TL tasks using public neuroimaging data for evaluation. Comprehensive experiments validate the superiority of SIDeR over ten competing methods, particularly an average improvement of 15.6% on the TL tasks with multi-source experiments.

No abstract

No abstract

Negative symptoms are a transdiagnostic feature of serious mental illness (SMI) that can be potentially “digitally phenotyped” using objective vocal analysis. In prior studies, vocal measures show low convergence with clinical ratings, potentially because analysis has used small, constrained acoustic feature sets. We sought to evaluate (1) whether clinically rated blunted vocal affect (BvA)/alogia could be accurately modelled using machine learning (ML) with a large feature set from two separate tasks (i.e., a 20-s “picture” and a 60-s “free-recall” task), (2) whether “Predicted” BvA/alogia (computed from the ML model) are associated with demographics, diagnosis, psychiatric symptoms, and cognitive/social functioning, and (3) which key vocal features are central to BvA/Alogia ratings. Accuracy was high (>90%) and was improved when computed separately by speaking task. ML scores were associated with poor cognitive performance and social functioning and were higher in patients with schizophrenia versus depression or mania diagnoses. However, the features identified as most predictive of BvA/Alogia were generally not considered critical to their operational definitions. Implications for validating and implementing digital phenotyping to reduce SMI burden are discussed.

No abstract