Bernhard Spitzer

Maintaining multiple items in working memory (WM) requires the allocation of limited resources, which may involve prioritizing individual items over others. Previous work has suggested that attentional prioritization can vary rhythmically between WM samples, but direct empirical evidence for such temporal fluctuations remains scarce. Here, we recorded eye-tracking and EEG to track the temporal dynamics of multi-item processing while participants maintained up to three object orientations in WM. Whereas EEG encoding was only short-lived, miniature gaze patterns reflected the orientations of all three items throughout the entire WM delay and showed systematic load and serial position effects. Contrary to rhythmic sampling accounts, however, we found no evidence that eye movements alternated between the concurrently maintained items. Furthermore, the strength of WM encoding in gaze was not modulated by frontal theta activity in EEG, suggesting that oculomotor encoding was not directly coupled to rhythmic neural dynamics associated with WM and internal attention. Together, our findings indicate that the oculomotor signature of visuospatial WM processing was rather stably shaped by all of the simultaneously maintained items. The results further corroborate that eye-tracking and EEG provide complementary windows into the temporal dynamics of WM.

Generalization enables individuals to apply knowledge derived from past experiences in novel situations, yet it is unclear whether common experimental paradigms probe the same underlying ability in early to middle childhood. We assessed 84 children aged 4, 6, and 8 years on four paradigms, namely statistical learning, associative inference, transitive inference, and categorization, and examined age effects and cross-task concordance. Accuracy increased with age in every task, but inter-task correlations were weak, suggesting that each paradigm might tap a partly distinct facet of generalization. Focused analyses of the associative and transitive inference task showed that older children’s success increasingly relied on memory for directly learned pairs. In younger children, the weak memory–generalization link diverged by task: 4- and 6-year-olds often solved transitive inferences without accurate pair memory, whereas such increased memory-independent generalization was not evident in associative inference. Computational modeling supported this pattern, indicating that younger children used value-based reinforcement learning to solve transitive problems, whereas 8-year-olds employed a pair-memory strategy. Taken together, our findings show that generalization develops through multiple pathways, underscoring the need for multi-indicator and model-based approaches to capture its multiple facets.

Anxiety has been linked to increased generalisation of threat expectations to perceptually similar stimuli. Such generalisation can arise either from a failure to distinguish threatening from non-threatening stimuli (perceptual mechanism) or from the transfer of learned values between stimuli (value-based mechanism). Yet, how these mechanisms contribute to generalisation remains unclear. Here we assess how participants ( n = 140) generalise outcome expectancies to perceptually similar stimuli, using personalised stimulus spaces. Computational modelling revealed that individuals differ in the extent to which they generalise value and in the underlying value function. We further found that stronger generalisation in trait anxiety was best explained by greater reliance on value transfer. In this work, we characterise individual differences in the generalisation of aversive stimuli and link stronger generalisation in trait anxiety to preferential reliance on value transfer.

Which factors determine whether information temporarily held in working memory (WM) can later be remembered from long-term memory (LTM)? Previous work has shown that retrieving (“testing”) memories from LTM can benefit their future LTM recall. Here, we examined the extent to which a benefit for subsequent LTM may also occur after retrieval from WM, depending on whether the WM contents were retrieved from a prioritized or deprioritized state. In three experiments ( n = 383 participants), we combined variants of a visual WM paradigm with a subsequent surprise LTM recall test. We found a LTM benefit of WM testing both for prioritized and deprioritized WM contents, which, interestingly, was stronger for the deprioritized information. This pattern showed similarly across experiments with different priority manipulations. Subsequent LTM benefits generally occurred after WM testing with a recall-like test format (continuous report), but not after simple WM comparisons against a probe. The surprisingly larger LTM benefit for deprioritized WM contents may reflect enhanced encoding of the participants’ own subjective WM report – as opposed to the originally presented sample information – into LTM.

Deep neural networks have achieved success across a wide range of applications, including as models of human behaviour and neural representations in vision tasks ^1,2 . However, neural network training and human learning differ in fundamental ways, and neural networks often fail to generalize as robustly as humans do ^3,4 , raising questions regarding the similarity of their underlying representations. We need to determine what is missing for modern learning systems to exhibit more human-aligned behaviour. Here we highlight a key misalignment between vision models and humans: whereas human conceptual knowledge is hierarchically organized from fine- to coarse-scale distinctions (for example, ref. ⁵ ), model representations do not accurately capture all these levels of abstraction. To address this misalignment, we first train a teacher model to imitate human judgements, then transfer human-aligned structure from its representations to refine the representations of pretrained state-of-the-art vision foundation models via fine-tuning. These human-aligned models more accurately approximate human behaviour and uncertainty across a wide range of similarity tasks, including a dataset of human judgements spanning multiple levels of semantic abstractions. They also perform better on a diverse set of machine learning tasks, increasing generalization and out-of-distribution robustness. Thus, infusing neural networks with additional human knowledge yields a best-of-both-worlds representation that is both more consistent with human cognitive judgements and more practically useful, paving the way towards more robust, interpretable and human-aligned artificial intelligence systems.

Which factors determine whether information temporarily held in working memory (WM) can later be remembered from long-term memory (LTM)? Previous work has shown that retrieving (“testing”) memories from LTM can benefit their future LTM recall. Here, we examined the extent to which a benefit for subsequent LTM may also occur after retrieval from WM, depending on whether the WM contents were retrieved from a prioritized or deprioritized state. In three experiments, we combined variants of a novel visual WM paradigm with a subsequent surprise LTM recall test. We found a LTM benefit of WM testing both for prioritized and deprioritized WM contents, which, interestingly, was stronger for the deprioritized information. This pattern showed similarly across experiments with different priority manipulations. Subsequent LTM benefits generally occurred after WM testing with a recall-like test format (continuous report), but not after simple WM comparisons against a probe. The surprisingly larger LTM benefit for deprioritized WM contents may reflect enhanced encoding of the participants’ own subjective WM report – as opposed to the originally presented sample information – into LTM.

People routinely make decisions based on samples of numerical values. A common conclusion from the literature in psychophysics and behavioral economics is that observers subjectively compress magnitudes, such that extreme values have less sway over people's decisions than prescribed by a normative model (underweighting). However, recent studies have reported evidence for anti-compression, that is, the relative overweighting of extreme values. Here, we investigate potential reasons for this discrepancy in findings and propose that it might reflect adaptive responses to different task requirements. We performed a large-scale study (n = 586) of sequential numerical integration, manipulating (a) the task requirement (averaging a single stream or comparing two interleaved streams of numbers), (b) the distribution of sample values (uniform or Gaussian), and (c) their range (1-9 or 100-900). The data showed compression of subjective values in the averaging task, but anticompression in the comparison task. This pattern held for both distribution types and for both ranges. In model simulations, we show that either compression or anticompression can be beneficial for noisy observers, depending on the sample-level processing demands imposed by the task. This suggests that the empirically observed patterns of over- and underweighting might reflect adaptive responses. (PsycInfo Database Record (c) 2024 APA, all rights reserved).

Stimulus-dependent eye movements have been recognized as a potential confound in decoding visual working memory information from neural signals. Here we combined eye-tracking with representational geometry analyses to uncover the information in miniature gaze patterns while participants (n = 41) were cued to maintain visual object orientations. Although participants were discouraged from breaking fixation by means of real-time feedback, small gaze shifts (<1°) robustly encoded the to-be-maintained stimulus orientation, with evidence for encoding two sequentially presented orientations at the same time. The orientation encoding on stimulus presentation was object-specific, but it changed to a more object-independent format during cued maintenance, particularly when attention had been temporarily withdrawn from the memorandum. Finally, categorical reporting biases increased after unattended storage, with indications of biased gaze geometries already emerging during the maintenance periods before behavioural reporting. These findings disclose a wealth of information in gaze patterns during visuospatial working memory and indicate systematic changes in representational format when memory contents have been unattended.

Stimulus-dependent eye movements have been recognized as a potential confound in decoding visual working memory information from neural signals. Here, we combined eye-tracking with representational geometry analyses to uncover the very information in miniature gaze patterns while participants (n = 41) were cued to maintain visual object orientations. Although participants were discouraged from breaking fixation via real-time feedback, small gaze shifts (< 1 degree) robustly encoded the to-be-maintained stimulus orientation, with evidence for encoding two sequentially presented orientations at the same time. While the orientation encoding upon stimulus presentation was object-specific, it changed to a more object-independent format during cued maintenance, particularly when attention had been temporarily withdrawn from the memorandum. Finally, categorical reporting biases increased after unattended storage, with indications of biased gaze geometries emerging already during the maintenance periods prior to behavioral reporting. These findings disclose a wealth of information in gaze patterns during visuospatial working memory, and suggest systematic changes in representational format when memory contents have been unattended.

Working Memory (WM) keeps information temporarily available for upcoming tasks. How the contents of WM are distinguished from perceptual representations on the one hand, and from long-term memories on the other, is still debated. Here, we leveraged recent evidence for a reversal of processing dynamics when retrieving episodic long-term memories as opposed to perceiving visual input. In two experiments (n=75 and n=103), we asked participants to hold one or more items in WM and to report their low-level perceptual and high-level semantic qualities. In both experiments, we found faster responses to the items’ semantic qualities, indicating prioritization over visuo-perceptual aspects, when two or more items were held concurrently in WM. These dynamics of accessing information in multi-item WM were akin to those in retrieving episodic long-term memories and opposite to those in processing visual inputs. Little to no semantic prioritization was evident during single-item maintenance, consistent with a strictly capacity-limited focus of attention within which WM information can be transformed into a prospective action plan.