The Geometry of Abstraction in the Hippocampus and Prefrontal Cortex.

The curse of dimensionality plagues models of reinforcement learning and decision making. The process of abstraction solves this by constructing variables describing features shared by different instances, reducing dimensionality and enabling generalization in novel situations. Here, we characterized neural representations in monkeys performing a task described by different hidden and explicit variables. Abstraction was defined operationally using the generalization performance of neural decoders across task conditions not used for training, which requires a particular geometry of neural representations. Neural ensembles in prefrontal cortex, hippocampus, and simulated neural networks simultaneously represented multiple variables in a geometry reflecting abstraction but that still allowed a linear classifier to decode a large number of other variables (high shattering dimensionality). Furthermore, this geometry changed in relation to task events and performance. These findings elucidate how the brain and artificial systems represent variables in an abstract format while preserving the advantages conferred by high shattering dimensionality.

Finding the Brain in the Nose.

Olfaction is fundamentally distinct from other sensory modalities. Natural odor stimuli are complex mixtures of volatile chemicals that interact in the nose with a receptor array that, in rodents, is built from more than 1,000 unique receptors. These interactions dictate a peripheral olfactory code, which in the brain is transformed and reformatted as it is broadcast across a set of highly interconnected olfactory regions. Here we discuss the problems of characterizing peripheral population codes for olfactory stimuli, of inferring the specific functions of different higher olfactory areas given their extensive recurrence, and of ultimately understanding how odor representations are linked to perception and action. We argue that, despite the differences between olfaction and other sensory modalities, addressing these specific questions will reveal general principles underlying brain function.

Peeling the Onion of Brain Representations.

The brain's function is to enable adaptive behavior in the world. To this end, the brain processes information about the world. The concept of representation links the information processed by the brain back to the world and enables us to understand what the brain does at a functional level. The appeal of making the connection between brain activity and what it represents has been irresistible to neuroscience, despite the fact that representational interpretations pose several challenges: We must define which aspects of brain activity matter, how the code works, and how it supports computations that contribute to adaptive behavior. It has been suggested that we might drop representational language altogether and seek to understand the brain, more simply, as a dynamical system. In this review, we argue that the concept of representation provides a useful link between dynamics and computational function and ask which aspects of brain activity should be analyzed to achieve a representational understanding. We peel the onion of brain representations in search of the layers (the aspects of brain activity) that matter to computation. The article provides an introduction to the motivation and mathematics of representational models, a critical discussion of their assumptions and limitations, and a preview of future directions in this area.

The time course of person perception from voices in the brain.

When listeners hear a voice, they rapidly form a complex first impression of who the person behind that voice might be. We characterize how these multivariate first impressions from voices emerge over time across different levels of abstraction using electroencephalography and representational similarity analysis. We find that for eight perceived physical (gender, age, and health), trait (attractiveness, dominance, and trustworthiness), and social characteristics (educatedness and professionalism), representations emerge early (~80 ms after stimulus onset), with voice acoustics contributing to those representations between ~100 ms and 400 ms. While impressions of person characteristics are highly correlated, we can find evidence for highly abstracted, independent representations of individual person characteristics. These abstracted representationse merge gradually over time. That is, representations of physical characteristics (age, gender) arise early (from ~120 ms), while representations of some trait and social characteristics emerge later (~360 ms onward). The findings align with recent theoretical models and shed light on the computations underpinning person perception from voices.

A mathematical theory of relational generalization in transitive inference.

Humans and animals routinely infer relations between different items or events and generalize these relations to novel combinations of items. This allows them to respond appropriately to radically novel circumstances and is fundamental to advanced cognition. However, how learning systems (including the brain) can implement the necessary inductive biases has been unclear. We investigated transitive inference (TI), a classic relational task paradigm in which subjects must learn a relation ([Formula: see text] and [Formula: see text]) and generalize it to new combinations of items ([Formula: see text]). Through mathematical analysis, we found that a broad range of biologically relevant learning models (e.g. gradient flow or ridge regression) perform TI successfully and recapitulate signature behavioral patterns long observed in living subjects. First, we found that models with item-wise additive representations automatically encode transitive relations. Second, for more general representations, a single scalar "conjunctivity factor" determines model behavior on TI and, further, the principle of norm minimization (a standard statistical inductive bias) enables models with fixed, partly conjunctive representations to generalize transitively. Finally, neural networks in the "rich regime," which enables representation learning and improves generalization on many tasks, unexpectedly show poor generalization and anomalous behavior on TI. We find that such networks implement a form of norm minimization (over hidden weights) that yields a local encoding mechanism lacking transitivity. Our findings show how minimal statistical learning principles give rise to a classical relational inductive bias (transitivity), explain empirically observed behaviors, and establish a formal approach to understanding the neural basis of relational abstraction.

Commutative avatars of representations of semisimple Lie groups.

Here we announce the construction and properties of a big commutative subalgebra of the Kirillov algebra attached to a finite dimensional irreducible representation of a complex semisimple Lie group. They are commutative finite flat algebras over the cohomology of the classifying space of the group. They are isomorphic with the equivariant intersection cohomology of affine Schubert varieties, endowing the latter with a new ring structure. Study of the finer aspects of the structure of the big algebras will also furnish the stalks of the intersection cohomology with ring structure, thus ringifying Lusztig's q-weight multiplicity polynomials i.e., certain affine Kazhdan-Lusztig polynomials.

Complementary multi-modality molecular self-supervised learning via non-overlapping masking for property prediction.

Self-supervised learning plays an important role in molecular representation learning because labeled molecular data are usually limited in many tasks, such as chemical property prediction and virtual screening. However, most existing molecular pre-training methods focus on one modality of molecular data, and the complementary information of two important modalities, SMILES and graph, is not fully explored. In this study, we propose an effective multi-modality self-supervised learning framework for molecular SMILES and graph. Specifically, SMILES data and graph data are first tokenized so that they can be processed by a unified Transformer-based backbone network, which is trained by a masked reconstruction strategy. In addition, we introduce a specialized non-overlapping masking strategy to encourage fine-grained interaction between these two modalities. Experimental results show that our framework achieves state-of-the-art performance in a series of molecular property prediction tasks, and a detailed ablation study demonstrates efficacy of the multi-modality framework and the masking strategy.

MvMRL: a multi-view molecular representation learning method for molecular property prediction.

Effective molecular representation learning is very important for Artificial Intelligence-driven Drug Design because it affects the accuracy and efficiency of molecular property prediction and other molecular modeling relevant tasks. However, previous molecular representation learning studies often suffer from limitations, such as over-reliance on a single molecular representation, failure to fully capture both local and global information in molecular structure, and ineffective integration of multiscale features from different molecular representations. These limitations restrict the complete and accurate representation of molecular structure and properties, ultimately impacting the accuracy of predicting molecular properties. To this end, we propose a novel multi-view molecular representation learning method called MvMRL, which can incorporate feature information from multiple molecular representations and capture both local and global information from different views well, thus improving molecular property prediction. Specifically, MvMRL consists of four parts: a multiscale CNN-SE Simplified Molecular Input Line Entry System (SMILES) learning component and a multiscale Graph Neural Network encoder to extract local feature information and global feature information from the SMILES view and the molecular graph view, respectively; a Multi-Layer Perceptron network to capture complex non-linear relationship features from the molecular fingerprint view; and a dual cross-attention component to fuse feature information on the multi-views deeply for predicting molecular properties. We evaluate the performance of MvMRL on 11 benchmark datasets, and experimental results show that MvMRL outperforms state-of-the-art methods, indicating its rationality and effectiveness in molecular property prediction. The source code of MvMRL was released in https://github.com/jedison-github/MvMRL.

Zariski density of crystalline points.

We show that crystalline points are Zariski dense in the deformation space of a representation of the absolute Galois group of a p-adic field. We also show that these points are dense in the subspace parameterizing deformations with the determinant equal to a fixed crystalline character. Our proof is purely local and works for all p-adic fields and all residual Galois representations.

The Geometry of Low- and High-Level Perceptual Spaces.

Low-level features are typically continuous (e.g., the gamut between two colors), but semantic information is often categorical (there is no corresponding gradient between dog and turtle) and hierarchical (animals live in land, water, or air). To determine the impact of these differences on cognitive representations, we characterized the geometry of perceptual spaces of five domains: a domain dominated by semantic information (animal names presented as words), a domain dominated by low-level features (colored textures), and three intermediate domains (animal images, lightly texturized animal images that were easy to recognize, and heavily texturized animal images that were difficult to recognize). Each domain had 37 stimuli derived from the same animal names. From 13 participants (9F), we gathered similarity judgments in each domain via an efficient psychophysical ranking paradigm. We then built geometric models of each domain for each participant, in which distances between stimuli accounted for participants' similarity judgments and intrinsic uncertainty. Remarkably, the five domains had similar global properties: each required 5-7 dimensions, and a modest amount of spherical curvature provided the best fit. However, the arrangement of the stimuli within these embeddings depended on the level of semantic information: dendrograms derived from semantic domains (word, image, and lightly texturized images) were more "tree-like" than those from feature-dominated domains (heavily texturized images and textures). Thus, the perceptual spaces of domains along this feature-dominated to semantic-dominated gradient shift to a tree-like organization when semantic information dominates, while retaining a similar global geometry.

Internal Representations Are Prioritized by Frontoparietal Theta Connectivity and Suppressed by alpha Oscillation Dynamics: Evidence from Concurrent Transcranial Magnetic Stimulation EEG and Invasive EEG.

Control over internal representations requires the prioritization of relevant information and suppression of irrelevant information. The frontoparietal network exhibits prominent neural oscillations during these distinct cognitive processes. Yet, the causal role of this network-scale activity is unclear. Here, we targeted theta-frequency frontoparietal coherence and dynamic alpha oscillations in the posterior parietal cortex using online rhythmic transcranial magnetic stimulation (TMS) in women and men while they prioritized or suppressed internally maintained working memory (WM) representations. Using concurrent high-density EEG, we provided evidence that we acutely drove the targeted neural oscillation and TMS improved WM capacity only when the evoked activity corresponded with the desired cognitive process. To suppress an internal representation, we increased the amplitude of lateralized alpha oscillations in the posterior parietal cortex contralateral to the irrelevant visual field. For prioritization, we found that TMS to the prefrontal cortex increased theta-frequency connectivity in the prefrontoparietal network contralateral to the relevant visual field. To understand the spatial specificity of these effects, we administered the WM task to participants with implanted electrodes. We found that theta connectivity during prioritization was directed from the lateral prefrontal to the superior posterior parietal cortex. Together, these findings provide causal evidence in support of a model where a frontoparietal theta network prioritizes internally maintained representations and alpha oscillations in the posterior parietal cortex suppress irrelevant representations.

FG-BERT: a generalized and self-supervised functional group-based molecular representation learning framework for properties prediction.

Artificial intelligence-based molecular property prediction plays a key role in molecular design such as bioactive molecules and functional materials. In this study, we propose a self-supervised pretraining deep learning (DL) framework, called functional group bidirectional encoder representations from transformers (FG-BERT), pertained based on ~1.45 million unlabeled drug-like molecules, to learn meaningful representation of molecules from function groups. The pretrained FG-BERT framework can be fine-tuned to predict molecular properties. Compared to state-of-the-art (SOTA) machine learning and DL methods, we demonstrate the high performance of FG-BERT in evaluating molecular properties in tasks involving physical chemistry, biophysics and physiology across 44 benchmark datasets. In addition, FG-BERT utilizes attention mechanisms to focus on FG features that are critical to the target properties, thereby providing excellent interpretability for downstream training tasks. Collectively, FG-BERT does not require any artificially crafted features as input and has excellent interpretability, providing an out-of-the-box framework for developing SOTA models for a variety of molecule (especially for drug) discovery tasks.

Short human eccDNAs are predictable from sequences.

BACKGROUND: Ubiquitous presence of short extrachromosomal circular DNAs (eccDNAs) in eukaryotic cells has perplexed generations of biologists. Their widespread origins in the genome lacking apparent specificity led some studies to conclude their formation as random or near-random. Despite this, the search for specific formation of short eccDNA continues with a recent surge of interest in biomarker development. RESULTS: To shed new light on the conflicting views on short eccDNAs' randomness, here we present DeepCircle, a bioinformatics framework incorporating convolution- and attention-based neural networks to assess their predictability. Short human eccDNAs from different datasets indeed have low similarity in genomic locations, but DeepCircle successfully learned shared DNA sequence features to make accurate cross-datasets predictions (accuracy: convolution-based models: 79.65 ± 4.7%, attention-based models: 83.31 ± 4.18%). CONCLUSIONS: The excellent performance of our models shows that the intrinsic predictability of eccDNAs is encoded in the sequences across tissue origins. Our work demonstrates how the perceived lack of specificity in genomics data can be re-assessed by deep learning models to uncover unexpected similarity.

Distinct neural mechanisms for action access and execution in the human brain: insights from an fMRI study.

Goal-directed actions are fundamental to human behavior, whereby inner goals are achieved through mapping action representations to motor outputs. The left premotor cortex (BA6) and the posterior portion of Broca's area (BA44) are two modulatory poles of the action system. However, how these regions support the representation-output mapping within the system is not yet understood. To address this, we conducted a finger-tapping functional magnetic resonance imaging experiment using action categories ranging from specific to general. Our study found distinct neural behaviors in BA44 and BA6 during action category processing and motor execution. During access of action categories, activity in a posterior portion of BA44 (pBA44) decreased linearly as action categories became less specific. Conversely, during motor execution, activity in BA6 increased linearly with less specific categories. These findings highlight the differential roles of pBA44 and BA6 in action processing. We suggest that pBA44 facilitates access to action categories by utilizing motor information from the behavioral context while the premotor cortex integrates motor information to execute the selected action. This finding enhances our understanding of the interplay between prefrontal cortical regions and premotor cortex in mapping action representation to motor execution and, more in general, of the cortical mechanisms underlying human behavior.

The neural representations of valence transformation in indole processing.

Indole is often associated with a sweet and floral odor typical of jasmine flowers at low concentrations and an unpleasant, animal-like odor at high concentrations. However, the mechanism whereby the brain processes this opposite valence of indole is not fully understood yet. In this study, we aimed to investigate the neural mechanisms underlying indole valence encoding in conversion and nonconversion groups using the smelling task to arouse pleasantness. For this purpose, 12 conversion individuals and 15 nonconversion individuals participated in an event-related functional magnetic resonance imaging paradigm with low (low-indole) and high (high-indole) indole concentrations in which valence was manipulated independent of intensity. The results of this experiment showed that neural activity in the right amygdala, orbitofrontal cortex and insula was associated with valence independent of intensity. Furthermore, activation in the right orbitofrontal cortex in response to low-indole was positively associated with subjective pleasantness ratings. Conversely, activation in the right insula and amygdala in response to low-indole was positively correlated with anticipatory hedonic traits. Interestingly, while amygdala activation in response to high-indole also showed a positive correlation with these hedonic traits, such correlation was observed solely with right insula activation in response to high-indole. Additionally, activation in the right amygdala in response to low-indole was positively correlated with consummatory pleasure and hedonic traits. Regarding olfactory function, only activation in the right orbitofrontal cortex in response to high-indole was positively correlated with olfactory identification, whereas activation in the insula in response to low-indole was negatively correlated with the level of self-reported olfactory dysfunction. Based on these findings, valence transformation of indole processing in the right orbitofrontal cortex, insula, and amygdala may be associated with individual hedonic traits and perceptual differences.

Neural computations in prosopagnosia.

We report an investigation of the neural processes involved in the processing of faces and objects of brain-lesioned patient PS, a well-documented case of pure acquired prosopagnosia. We gathered a substantial dataset of high-density electrophysiological recordings from both PS and neurotypicals. Using representational similarity analysis, we produced time-resolved brain representations in a format that facilitates direct comparisons across time points, different individuals, and computational models. To understand how the lesions in PS's ventral stream affect the temporal evolution of her brain representations, we computed the temporal generalization of her brain representations. We uncovered that PS's early brain representations exhibit an unusual similarity to later representations, implying an excessive generalization of early visual patterns. To reveal the underlying computational deficits, we correlated PS' brain representations with those of deep neural networks (DNN). We found that the computations underlying PS' brain activity bore a closer resemblance to early layers of a visual DNN than those of controls. However, the brain representations in neurotypicals became more akin to those of the later layers of the model compared to PS. We confirmed PS's deficits in high-level brain representations by demonstrating that her brain representations exhibited less similarity with those of a DNN of semantics.

Similar patterns of brain activity for holistic representation of working memory: a study in children and adults.

Working memory is the fundamental function of the various cognitive processes and abilities in the overall trajectory of development. Significant advances in multivariate analysis of human functional magnetic resonance imaging data have converged functional segregation models toward integrated representation-based models. However, due to the inherent limitations of the multi-voxel pattern analysis method, we are unable to determine whether the underlying neural representations are spatially similar in the brain. Our study attempts to answer this question by examining the spatial similarity of brain activity during the working memory task in children and adults. Our results reveal similar patterns of activity between the regions involved in working memory. This functional network of similar spatial patterns was observed in both normally developing children and adults. However, the between-region similarity was more pronounced in adults than in children and associated with better performance. We propose an exchange of similar information flows through the brain at an integrated level of working memory processes, underpinning the holistic nature of working memory representation.

Multiple and Dissociable Effects of Sensory History on Working-Memory Performance.

Behavioral reports of sensory information are biased by stimulus history. The nature and direction of such serial-dependence biases can differ between experimental settings; both attractive and repulsive biases toward previous stimuli have been observed. How and when these biases arise in the human brain remains largely unexplored. They could occur either via a change in sensory processing itself and/or during postperceptual processes such as maintenance or decision-making. To address this, we tested 20 participants (11 female) and analyzed behavioral and magnetoencephalographic (MEG) data from a working-memory task in which participants were sequentially presented with two randomly oriented gratings, one of which was cued for recall at the end of the trial. Behavioral responses showed evidence for two distinct biases: (1) a within-trial repulsive bias away from the previously encoded orientation on the same trial, and (2) a between-trial attractive bias toward the task-relevant orientation on the previous trial. Multivariate classification of stimulus orientation revealed that neural representations during stimulus encoding were biased away from the previous grating orientation, regardless of whether we considered the within-trial or between-trial prior orientation, despite opposite effects on behavior. These results suggest that repulsive biases occur at the level of sensory processing and can be overridden at postperceptual stages to result in attractive biases in behavior.SIGNIFICANCE STATEMENT Recent experience biases behavioral reports of sensory information, possibly capitalizing on the temporal regularity in our environment. It is still unclear at what stage of stimulus processing such serial biases arise. Here, we recorded behavior and neurophysiological [magnetoencephalographic (MEG)] data to test whether neural activity patterns during early sensory processing show the same biases seen in participants' reports. In a working-memory task that produced multiple biases in behavior, responses were biased toward previous targets, but away from more recent stimuli. Neural activity patterns were uniformly biased away from all previously relevant items. Our results contradict proposals that all serial biases arise at an early sensory processing stage. Instead, neural activity exhibited mostly adaptation-like responses to recent stimuli.

Predicting and coding sound into action translation in spinal cord injured people.

Motor activation in response to perception of action-related stimuli may depend on a resonance mechanism subserving action understanding. The extent to which this mechanism is innate or learned from sensorimotor experience is still unclear. Here, we recorded EEG while people with paraplegia or tetraplegia consequent to spinal cord injury (SCI) and healthy control participants were presented with action sounds produced by body parts (mouth, hands or feet) that were or were not affected by SCI. Non-action sounds were used as further control. We observed reduced brain activation in subjects affected by SCI at both pre- and post-stimulus latencies specifically for those actions whose effector was disconnected by the spinal lesion (i.e., hand sound for tetraplegia and leg sound for both paraplegia and tetraplegia). Correlation analyses showed that these modulations were functionally linked with the chronicity of the lesion, indicating that the longer the time the lesion- EEG data acquisition interval and/or the more the lesion occurred at a young age, the weaker was the cortical activity in response to these action sounds. Tellingly, source estimations confirmed that these modulations originated from a deficit in the motor resonance mechanism, by showing diminished activity in premotor (during prediction and perception) and near the primary motor (during perception) areas. Such dissociation along the cortical hierarchy is consistent with both previous reports in healthy subjects and with hierarchical predictive coding accounts. Overall, these data expand on the notion that sensorimotor experience maintains the cortical representations relevant to anticipate and perceive action-related stimuli.

Individual word representations dissociate from linguistic context along a cortical unimodal to heteromodal gradient.

Language comprehension involves multiple hierarchical processing stages across time, space, and levels of representation. When processing a word, the sensory input is transformed into increasingly abstract representations that need to be integrated with the linguistic context. Thus, language comprehension involves both input-driven as well as context-dependent processes. While neuroimaging research has traditionally focused on mapping individual brain regions to the distinct underlying processes, recent studies indicate that whole-brain distributed patterns of cortical activation might be highly relevant for cognitive functions, including language. One such pattern, based on resting-state connectivity, is the 'principal cortical gradient', which dissociates sensory from heteromodal brain regions. The present study investigated the extent to which this gradient provides an organizational principle underlying language function, using a multimodal neuroimaging dataset of functional magnetic resonance imaging (fMRI) and magnetoencephalography (MEG) recordings from 102 participants during sentence reading. We found that the brain response to individual representations of a word (word length, orthographic distance, and word frequency), which reflect visual; orthographic; and lexical properties, gradually increases towards the sensory end of the gradient. Although these properties showed opposite effect directions in fMRI and MEG, their association with the sensory end of the gradient was consistent across both neuroimaging modalities. In contrast, MEG revealed that properties reflecting a word's relation to its linguistic context (semantic similarity and position within the sentence) involve the heteromodal end of the gradient to a stronger extent. This dissociation between individual word and contextual properties was stable across earlier and later time windows during word presentation, indicating interactive processing of word representations and linguistic context at opposing ends of the principal gradient. To conclude, our findings indicate that the principal gradient underlies the organization of a range of linguistic representations while supporting a gradual distinction between context-independent and context-dependent representations. Furthermore, the gradient reveals convergent patterns across neuroimaging modalities (similar location along the gradient) in the presence of divergent responses (opposite effect directions).