Neural representational geometry underlies few-shot concept learning

The cortical critical power law balances energy and information in an optimal fashion

A recent study has suggested that the stimulus responses of cortical neural populations follow a critical power law. More precisely, the power spectrum of the covariance matrix of neural responses follows a power law with an exponent indicating that the neural manifold lies on the edge of differentiability. This criticality is hypothesized to balance expressivity and robustness in neural encoding, as population responses on a nondifferential fractal manifold are thought to be overly sensitive to perturbations. However, contrary to this hypothesis, we prove that neural coding is far more robust than previously assumed. We develop a theoretical framework that provides an analytical expression for the Fisher information of population coding under the small noise assumption. Our results reveal that, due to its intrinsic high dimensionality, population coding maintains reliability even on a nondifferentiable fractal manifold, despite its sensitivity to perturbations. Furthermore, the theory reveals that the trade-off between energetic cost and information makes the critical power-law coding the optimal neural encoding of sensory information for a wide range of conditions. In this derivation, we highlight the essential role of a neural correlation, known as differential correlation, in power-law population coding. By uncovering the nontrivial nature of high-dimensional information coding, this work deepens our understanding of criticality and power laws in both biological and artificial neural computation.

Coding schemes in neural networks learning classification tasks

Neural networks posses the crucial ability to generate meaningful representations of task-dependent features. Indeed, with appropriate scaling, supervised learning in neural networks can result in strong, task-dependent feature learning. However, the nature of the emergent representations is still unclear. To understand the effect of learning on representations, we investigate fully-connected, wide neural networks learning classification tasks using the Bayesian framework where learning shapes the posterior distribution of the network weights. Consistent with previous findings, our analysis of the feature learning regime (also known as 'non-lazy' regime) shows that the networks acquire strong, data-dependent features, denoted as coding schemes, where neuronal responses to each input are dominated by its class membership. Surprisingly, the nature of the coding schemes depends crucially on the neuronal nonlinearity. In linear networks, an analog coding scheme of the task emerges; in nonlinear networks, strong spontaneous symmetry breaking leads to either redundant or sparse coding schemes. Our findings highlight how network properties such as scaling of weights and neuronal nonlinearity can profoundly influence the emergent representations.

Cell-type-specific manifold analysis discloses independent geometric transformations in the hippocampal spatial code

Integrating analyses of genetically defined cell types with population-level approaches remains poorly explored. We investigated this question by focusing on hippocampal spatial maps and the contribution of two genetically defined pyramidal cell types in the deep and superficial CA1 sublayers. Using single- and dual-color miniscope imaging in mice running along a linear track, we found that population activity from these cells exhibited three-dimensional ring manifolds that encoded the animal position and running direction. Despite shared topology, sublayer-specific manifolds displayed distinct geometric features. Manipulating track orientation revealed rotational and translational changes in manifolds from deep cells, contrasting with more stable representations by superficial cells. These transformations were not observed in manifolds derived from the entire CA1 population. Instead, cell-type-specific chemogenetic silencing of either sublayer revealed independent geometric codes. Our results show how genetically specified subpopulations may underpin parallel spatial maps that can be manipulated independently.

Nonlinear classification of neural manifolds with contextual information

Understanding how neural systems efficiently process information through distributed representations is a fundamental challenge at the interface of neuroscience and machine learning. Recent approaches analyze the statistical and geometrical attributes of neural representations as population-level mechanistic descriptors of task implementation. In particular, manifold capacity has emerged as a promising framework linking population geometry to the separability of neural manifolds. However, this metric has been limited to linear readouts. To address this limitation, we introduce a theoretical framework that leverages latent directions in input space, which can be related to contextual information. We derive an exact formula for the context-dependent manifold capacity that depends on manifold geometry and context correlations, and validate it on synthetic and real data. Our framework's increased expressivity captures representation reformatting in deep networks at early stages of the layer hierarchy, previously inaccessible to analysis. As context-dependent nonlinearity is ubiquitous in neural systems, our data-driven and theoretically grounded approach promises to elucidate context-dependent computation across scales, datasets, and models.

Unraveling the Geometry of Visual Relational Reasoning

Humans and other animals readily generalize abstract relations, such as recognizing constant in shape or color, whereas neural networks struggle. To investigate how neural networks generalize abstract relations, we introduce SimplifiedRPM, a novel benchmark for systematic evaluation. In parallel, we conduct human experiments to benchmark relational difficulty, enabling direct model-human comparisons. Testing four architectures-ResNet-50, Vision Transformer, Wild Relation Network, and Scattering Compositional Learner (SCL)-we find that SCL best aligns with human behavior and generalizes best. Building on a geometric theory of neural representations, we show representational geometries that predict generalization. Layer-wise analysis reveals distinct relational reasoning strategies across models and suggests a trade-off where unseen rule representations compress into training-shaped subspaces. Guided by our geometric perspective, we propose and evaluate SNRloss, a novel objective balancing representation geometry. Our findings offer geometric insights into how neural networks generalize abstract relations, paving the way for more human-like visual reasoning in AI.

Bridging Neuroscience and AI: Environmental Enrichment as a model for forward knowledge transfer in continual learning

Continual learning (CL) refers to an agent's capability to learn from a continuous stream of data and transfer knowledge without forgetting old information. One crucial aspect of CL is forward transfer, i.e., improved and faster learning on a new task by leveraging information from prior knowledge. While this ability comes naturally to biological brains, it poses a significant challenge for artificial intelligence (AI). Here, we suggest that environmental enrichment (EE) can be used as a biological model for studying forward transfer, inspiring human-like AI development. EE refers to animal studies that enhance cognitive, social, motor, and sensory stimulation and is a model for what, in humans, is referred to as 'cognitive reserve'. Enriched animals show significant improvement in learning speed and performance on new tasks, typically exhibiting forward transfer. We explore anatomical, molecular, and neuronal changes post-EE and discuss how artificial neural networks (ANNs) can be used to predict neural computation changes after enriched experiences. Finally, we provide a synergistic way of combining neuroscience and AI research that paves the path toward developing AI capable of rapid and efficient new task learning.

Informational ecosystems partially explain differences in socioenvironmental conceptual associations between U.S. American racial groups

Social groups represent a collective identity defined by a distinct consensus of concepts (e.g., ideas, values, and goals) whose structural relationship varies between groups. Here we set out to measure how a set of inter-concept semantic associations, comprising what we refer to as a concept graph, covaries between established social groups, based on racial identity, and how this effect is mediated by information ecosystems, contextualized as news sources. Group differences among racial identity (278 Black and 294 white Americans) and informational ecosystems (Left- and Right- leaning news sources) are present in subjective judgments of how the meaning of concepts such as healthcare, police, and voting relate to each other. These racial group differences in concept graphs were partially mediated by the bias of news sources that individuals get their information from. This supports the idea of groups being defined by common conceptual semantic relationships that partially arise from shared information ecosystems.

Semi-orthogonal subspaces for value mediate a binding and generalization trade-off

When choosing between options, we must associate their values with the actions needed to select them. We hypothesize that the brain solves this binding problem through neural population subspaces. Here, in macaques performing a choice task, we show that neural populations in five reward-sensitive regions encode the values of offers presented on the left and right in distinct subspaces. This encoding is sufficient to bind offer values to their locations while preserving abstract value information. After offer presentation, all areas encode the value of the first and second offers in orthogonal subspaces; this orthogonalization also affords binding. Our binding-by-subspace hypothesis makes two new predictions confirmed by the data. First, behavioral errors should correlate with spatial, but not temporal, neural misbinding. Second, behavioral errors should increase when offers have low or high values, compared to medium values, even when controlling for value difference. Together, these results support the idea that the brain uses semi-orthogonal subspaces to bind features.

A large-scale examination of inductive biases shaping high-level visual representation in brains and machines

The rapid release of high-performing computer vision models offers new potential to study the impact of different inductive biases on the emergent brain alignment of learned representations. Here, we perform controlled comparisons among a curated set of 224 diverse models to test the impact of specific model properties on visual brain predictivity - a process requiring over 1.8 billion regressions and 50.3 thousand representational similarity analyses. We find that models with qualitatively different architectures (e.g. CNNs versus Transformers) and task objectives (e.g. purely visual contrastive learning versus vision- language alignment) achieve near equivalent brain predictivity, when other factors are held constant. Instead, variation across visual training diets yields the largest, most consistent effect on brain predictivity. Many models achieve similarly high brain predictivity, despite clear variation in their underlying representations - suggesting that standard methods used to link models to brains may be too flexible. Broadly, these findings challenge common assumptions about the factors underlying emergent brain alignment, and outline how we can leverage controlled model comparison to probe the common computational principles underlying biological and artificial visual systems.

The topology and geometry of neural representations

A central question for neuroscience is how to characterize brain representations of perceptual and cognitive content. An ideal characterization should distinguish different functional regions with robustness to noise and idiosyncrasies of individual brains that do not correspond to computational differences. Previous studies have characterized brain representations by their representational geometry, which is defined by the representational dissimilarity matrix (RDM), a summary statistic that abstracts from the roles of individual neurons (or responses channels) and characterizes the discriminability of stimuli. Here, we explore a further step of abstraction: from the geometry to the topology of brain representations. We propose topological representational similarity analysis, an extension of representational similarity analysis that uses a family of geotopological summary statistics that generalizes the RDM to characterize the topology while de-emphasizing the geometry. We evaluate this family of statistics in terms of the sensitivity and specificity for model selection using both simulations and functional MRI (fMRI) data. In the simulations, the ground truth is a data-generating layer representation in a neural network model and the models are the same and other layers in different model instances (trained from different random seeds). In fMRI, the ground truth is a visual area and the models are the same and other areas measured in different subjects. Results show that topology-sensitive characterizations of population codes are robust to noise and interindividual variability and maintain excellent sensitivity to the unique representational signatures of different neural network layers and brain regions.

EEG spectral attractors identify a geometric core of brain dynamics

Multidimensional reconstruction of brain attractors from electroencephalography (EEG) data enables the analysis of geometric complexity and interactions between signals in state space. Utilizing resting-state data from young and older adults, we characterize periodic (traditional frequency bands) and aperiodic (broadband exponent) attractors according to their geometric complexity and shared dynamical signatures, which we refer to as a geometric cross-parameter coupling. Alpha and aperiodic attractors are the least complex, and their global shapes are shared among all other frequency bands, affording alpha and aperiodic greater predictive power. Older adults show lower geometric complexity but greater coupling, resulting from dedifferentiation of gamma activity. The form and content of resting-state thoughts were further associated with the complexity of attractor dynamics. These findings support a process-developmental perspective on the brain's dynamic core, whereby more complex information differentiates out of an integrative and global geometric core.

Factorized visual representations in the primate visual system and deep neural networks

Object classification has been proposed as a principal objective of the primate ventral visual stream and has been used as an optimization target for deep neural network models (DNNs) of the visual system. However, visual brain areas represent many different types of information, and optimizing for classification of object identity alone does not constrain how other information may be encoded in visual representations. Information about different scene parameters may be discarded altogether ('invariance'), represented in non-interfering subspaces of population activity ('factorization') or encoded in an entangled fashion. In this work, we provide evidence that factorization is a normative principle of biological visual representations. In the monkey ventral visual hierarchy, we found that factorization of object pose and background information from object identity increased in higher-level regions and strongly contributed to improving object identity decoding performance. We then conducted a large-scale analysis of factorization of individual scene parameters - lighting, background, camera viewpoint, and object pose - in a diverse library of DNN models of the visual system. Models which best matched neural, fMRI, and behavioral data from both monkeys and humans across 12 datasets tended to be those which factorized scene parameters most strongly. Notably, invariance to these parameters was not as consistently associated with matches to neural and behavioral data, suggesting that maintaining non-class information in factorized activity subspaces is often preferred to dropping it altogether. Thus, we propose that factorization of visual scene information is a widely used strategy in brains and DNN models thereof.

Representations and generalization in artificial and brain neural networks

Humans and animals excel at generalizing from limited data, a capability yet to be fully replicated in artificial intelligence. This perspective investigates generalization in biological and artificial deep neural networks (DNNs), in both in-distribution and out-of-distribution contexts. We introduce two hypotheses: First, the geometric properties of the neural manifolds associated with discrete cognitive entities, such as objects, words, and concepts, are powerful order parameters. They link the neural substrate to the generalization capabilities and provide a unified methodology bridging gaps between neuroscience, machine learning, and cognitive science. We overview recent progress in studying the geometry of neural manifolds, particularly in visual object recognition, and discuss theories connecting manifold dimension and radius to generalization capacity. Second, we suggest that the theory of learning in wide DNNs, especially in the thermodynamic limit, provides mechanistic insights into the learning processes generating desired neural representational geometries and generalization. This includes the role of weight norm regularization, network architecture, and hyper-parameters. We will explore recent advances in this theory and ongoing challenges. We also discuss the dynamics of learning and its relevance to the issue of representational drift in the brain.

Alignment of brain embeddings and artificial contextual embeddings in natural language points to common geometric patterns

Contextual embeddings, derived from deep language models (DLMs), provide a continuous vectorial representation of language. This embedding space differs fundamentally from the symbolic representations posited by traditional psycholinguistics. We hypothesize that language areas in the human brain, similar to DLMs, rely on a continuous embedding space to represent language. To test this hypothesis, we densely record the neural activity patterns in the inferior frontal gyrus (IFG) of three participants using dense intracranial arrays while they listened to a 30-minute podcast. From these fine-grained spatiotemporal neural recordings, we derive a continuous vectorial representation for each word (i.e., a brain embedding) in each patient. Using stringent zero-shot mapping we demonstrate that brain embeddings in the IFG and the DLM contextual embedding space have common geometric patterns. The common geometric patterns allow us to predict the brain embedding in IFG of a given left-out word based solely on its geometrical relationship to other non-overlapping words in the podcast. Furthermore, we show that contextual embeddings capture the geometry of IFG embeddings better than static word embeddings. The continuous brain embedding space exposes a vector-based neural code for natural language processing in the human brain.

Population encoding of stimulus features along the visual hierarchy

The retina and primary visual cortex (V1) both exhibit diverse neural populations sensitive to diverse visual features. Yet it remains unclear how neural populations in each area partition stimulus space to span these features. One possibility is that neural populations are organized into discrete groups of neurons, with each group signaling a particular constellation of features. Alternatively, neurons could be continuously distributed across feature-encoding space. To distinguish these possibilities, we presented a battery of visual stimuli to the mouse retina and V1 while measuring neural responses with multi-electrode arrays. Using machine learning approaches, we developed a manifold embedding technique that captures how neural populations partition feature space and how visual responses correlate with physiological and anatomical properties of individual neurons. We show that retinal populations discretely encode features, while V1 populations provide a more continuous representation. Applying the same analysis approach to convolutional neural networks that model visual processing, we demonstrate that they partition features much more similarly to the retina, indicating they are more like big retinas than little brains.

Scientists Invent New Hypotheses, Do Brains?

How are new Bayesian hypotheses generated within the framework of predictive processing? This explanatory framework purports to provide a unified, systematic explanation of cognition by appealing to Bayes rule and hierarchical Bayesian machinery alone. Given that the generation of new hypotheses is fundamental to Bayesian inference, the predictive processing framework faces an important challenge in this regard. By examining several cognitive-level and neurobiological architecture-inspired models of hypothesis generation, we argue that there is an essential difference between the two types of models. Cognitive-level models do not specify how they can be implemented in brains and include structures and assumptions that are external to the predictive processing framework. By contrast, neurobiological architecture-inspired models, which aim to better resemble brain processes, fail to explain important capacities of cognition, such as categorization and few-shot learning. The "scaling-up" challenge for proponents of predictive processing is to explain the relationship between these two types of models using only the theoretical and conceptual machinery of Bayesian inference.

High-performing neural network models of visual cortex benefit from high latent dimensionality

Geometric descriptions of deep neural networks (DNNs) have the potential to uncover core representational principles of computational models in neuroscience. Here we examined the geometry of DNN models of visual cortex by quantifying the latent dimensionality of their natural image representations. A popular view holds that optimal DNNs compress their representations onto low-dimensional subspaces to achieve invariance and robustness, which suggests that better models of visual cortex should have lower dimensional geometries. Surprisingly, we found a strong trend in the opposite direction-neural networks with high-dimensional image subspaces tended to have better generalization performance when predicting cortical responses to held-out stimuli in both monkey electrophysiology and human fMRI data. Moreover, we found that high dimensionality was associated with better performance when learning new categories of stimuli, suggesting that higher dimensional representations are better suited to generalize beyond their training domains. These findings suggest a general principle whereby high-dimensional geometry confers computational benefits to DNN models of visual cortex.

A model of conceptual bootstrapping in human cognition

To tackle a hard problem, it is often wise to reuse and recombine existing knowledge. Such an ability to bootstrap enables us to grow rich mental concepts despite limited cognitive resources. Here we present a computational model of conceptual bootstrapping. This model uses a dynamic conceptual repertoire that can cache and later reuse elements of earlier insights in principled ways, modelling learning as a series of compositional generalizations. This model predicts systematically different learned concepts when the same evidence is processed in different orders, without any extra assumptions about previous beliefs or background knowledge. Across four behavioural experiments (total n = 570), we demonstrate strong curriculum-order and conceptual garden-pathing effects that closely resemble our model predictions and differ from those of alternative accounts. Taken together, this work offers a computational account of how past experiences shape future conceptual discoveries and showcases the importance of curriculum design in human inductive concept inferences.

How well do rudimentary plasticity rules predict adult visual object learning?

A core problem in visual object learning is using a finite number of images of a new object to accurately identify that object in future, novel images. One longstanding, conceptual hypothesis asserts that this core problem is solved by adult brains through two connected mechanisms: 1) the re-representation of incoming retinal images as points in a fixed, multidimensional neural space, and 2) the optimization of linear decision boundaries in that space, via simple plasticity rules applied to a single downstream layer. Though this scheme is biologically plausible, the extent to which it explains learning behavior in humans has been unclear-in part because of a historical lack of image-computable models of the putative neural space, and in part because of a lack of measurements of human learning behaviors in difficult, naturalistic settings. Here, we addressed these gaps by 1) drawing from contemporary, image-computable models of the primate ventral visual stream to create a large set of testable learning models (n = 2,408 models), and 2) using online psychophysics to measure human learning trajectories over a varied set of tasks involving novel 3D objects (n = 371,000 trials), which we then used to develop (and publicly release) empirical benchmarks for comparing learning models to humans. We evaluated each learning model on these benchmarks, and found those based on deep, high-level representations from neural networks were surprisingly aligned with human behavior. While no tested model explained the entirety of replicable human behavior, these results establish that rudimentary plasticity rules, when combined with appropriate visual representations, have high explanatory power in predicting human behavior with respect to this core object learning problem.

Exploring strategy differences between humans and monkeys with recurrent neural networks

Animal models are used to understand principles of human biology. Within cognitive neuroscience, non-human primates are considered the premier model for studying decision-making behaviors in which direct manipulation experiments are still possible. Some prominent studies have brought to light major discrepancies between monkey and human cognition, highlighting problems with unverified extrapolation from monkey to human. Here, we use a parallel model system-artificial neural networks (ANNs)-to investigate a well-established discrepancy identified between monkeys and humans with a working memory task, in which monkeys appear to use a recency-based strategy while humans use a target-selective strategy. We find that ANNs trained on the same task exhibit a progression of behavior from random behavior (untrained) to recency-like behavior (partially trained) and finally to selective behavior (further trained), suggesting monkeys and humans may occupy different points in the same overall learning progression. Surprisingly, what appears to be recency-like behavior in the ANN, is in fact an emergent non-recency-based property of the organization of the neural network's state space during its development through training. We find that explicit encouragement of recency behavior during training has a dual effect, not only causing an accentuated recency-like behavior, but also speeding up the learning process altogether, resulting in an efficient shaping mechanism to achieve the optimal strategy. Our results suggest a new explanation for the discrepency observed between monkeys and humans and reveal that what can appear to be a recency-based strategy in some cases may not be recency at all.

Dimension of Activity in Random Neural Networks

Neural networks are high-dimensional nonlinear dynamical systems that process information through the coordinated activity of many connected units. Understanding how biological and machine-learning networks function and learn requires knowledge of the structure of this coordinated activity, information contained, for example, in cross covariances between units. Self-consistent dynamical mean field theory (DMFT) has elucidated several features of random neural networks-in particular, that they can generate chaotic activity-however, a calculation of cross covariances using this approach has not been provided. Here, we calculate cross covariances self-consistently via a two-site cavity DMFT. We use this theory to probe spatiotemporal features of activity coordination in a classic random-network model with independent and identically distributed (i.i.d.) couplings, showing an extensive but fractionally low effective dimension of activity and a long population-level timescale. Our formulas apply to a wide range of single-unit dynamics and generalize to non-i.i.d. couplings. As an example of the latter, we analyze the case of partially symmetric couplings.

Semi-orthogonal subspaces for value mediate a tradeoff between binding and generalization

When choosing between options, we must associate their values with the action needed to select them. We hypothesize that the brain solves this binding problem through neural population subspaces. To test this hypothesis, we examined neuronal responses in five reward-sensitive regions in macaques performing a risky choice task with sequential offers. Surprisingly, in all areas, the neural population encoded the values of offers presented on the left and right in distinct subspaces. We show that the encoding we observe is sufficient to bind the values of the offers to their respective positions in space while preserving abstract value information, which may be important for rapid learning and generalization to novel contexts. Moreover, after both offers have been presented, all areas encode the value of the first and second offers in orthogonal subspaces. In this case as well, the orthogonalization provides binding. Our binding-by-subspace hypothesis makes two novel predictions borne out by the data. First, behavioral errors should correlate with putative spatial (but not temporal) misbinding in the neural representation. Second, the specific representational geometry that we observe across animals also indicates that behavioral errors should increase when offers have low or high values, compared to when they have medium values, even when controlling for value difference. Together, these results support the idea that the brain makes use of semi-orthogonal subspaces to bind features together.

Dimensionality and Ramping: Signatures of Sentence Integration in the Dynamics of Brains and Deep Language Models

A sentence is more than the sum of its words: its meaning depends on how they combine with one another. The brain mechanisms underlying such semantic composition remain poorly understood. To shed light on the neural vector code underlying semantic composition, we introduce two hypotheses: (1) the intrinsic dimensionality of the space of neural representations should increase as a sentence unfolds, paralleling the growing complexity of its semantic representation; and (2) this progressive integration should be reflected in ramping and sentence-final signals. To test these predictions, we designed a dataset of closely matched normal and jabberwocky sentences (composed of meaningless pseudo words) and displayed them to deep language models and to 11 human participants (5 men and 6 women) monitored with simultaneous MEG and intracranial EEG. In both deep language models and electrophysiological data, we found that representational dimensionality was higher for meaningful sentences than jabberwocky. Furthermore, multivariate decoding of normal versus jabberwocky confirmed three dynamic patterns: (1) a phasic pattern following each word, peaking in temporal and parietal areas; (2) a ramping pattern, characteristic of bilateral inferior and middle frontal gyri; and (3) a sentence-final pattern in left superior frontal gyrus and right orbitofrontal cortex. These results provide a first glimpse into the neural geometry of semantic integration and constrain the search for a neural code of linguistic composition.SIGNIFICANCE STATEMENT Starting from general linguistic concepts, we make two sets of predictions in neural signals evoked by reading multiword sentences. First, the intrinsic dimensionality of the representation should grow with additional meaningful words. Second, the neural dynamics should exhibit signatures of encoding, maintaining, and resolving semantic composition. We successfully validated these hypotheses in deep neural language models, artificial neural networks trained on text and performing very well on many natural language processing tasks. Then, using a unique combination of MEG and intracranial electrodes, we recorded high-resolution brain data from human participants while they read a controlled set of sentences. Time-resolved dimensionality analysis showed increasing dimensionality with meaning, and multivariate decoding allowed us to isolate the three dynamical patterns we had hypothesized.

Linear Classification of Neural Manifolds with Correlated Variability

Understanding how the statistical and geometric properties of neural activity relate to performance is a key problem in theoretical neuroscience and deep learning. Here, we calculate how correlations between object representations affect the capacity, a measure of linear separability. We show that for spherical object manifolds, introducing correlations between centroids effectively pushes the spheres closer together, while introducing correlations between the axes effectively shrinks their radii, revealing a duality between correlations and geometry with respect to the problem of classification. We then apply our results to accurately estimate the capacity of deep network data.

Manifold learning for fMRI time-varying functional connectivity

Whole-brain functional connectivity (FC) measured with functional MRI (fMRI) evolves over time in meaningful ways at temporal scales going from years (e.g., development) to seconds [e.g., within-scan time-varying FC (tvFC)]. Yet, our ability to explore tvFC is severely constrained by its large dimensionality (several thousands). To overcome this difficulty, researchers often seek to generate low dimensional representations (e.g., 2D and 3D scatter plots) hoping those will retain important aspects of the data (e.g., relationships to behavior and disease progression). Limited prior empirical work suggests that manifold learning techniques (MLTs)-namely those seeking to infer a low dimensional non-linear surface (i.e., the manifold) where most of the data lies-are good candidates for accomplishing this task. Here we explore this possibility in detail. First, we discuss why one should expect tvFC data to lie on a low dimensional manifold. Second, we estimate what is the intrinsic dimension (ID; i.e., minimum number of latent dimensions) of tvFC data manifolds. Third, we describe the inner workings of three state-of-the-art MLTs: Laplacian Eigenmaps (LEs), T-distributed Stochastic Neighbor Embedding (T-SNE), and Uniform Manifold Approximation and Projection (UMAP). For each method, we empirically evaluate its ability to generate neuro-biologically meaningful representations of tvFC data, as well as their robustness against hyper-parameter selection. Our results show that tvFC data has an ID that ranges between 4 and 26, and that ID varies significantly between rest and task states. We also show how all three methods can effectively capture subject identity and task being performed: UMAP and T-SNE can capture these two levels of detail concurrently, but LE could only capture one at a time. We observed substantial variability in embedding quality across MLTs, and within-MLT as a function of hyper-parameter selection. To help alleviate this issue, we provide heuristics that can inform future studies. Finally, we also demonstrate the importance of feature normalization when combining data across subjects and the role that temporal autocorrelation plays in the application of MLTs to tvFC data. Overall, we conclude that while MLTs can be useful to generate summary views of labeled tvFC data, their application to unlabeled data such as resting-state remains challenging.

Mechanisms of human dynamic object recognition revealed by sequential deep neural networks

Humans can quickly recognize objects in a dynamically changing world. This ability is showcased by the fact that observers succeed at recognizing objects in rapidly changing image sequences, at up to 13 ms/image. To date, the mechanisms that govern dynamic object recognition remain poorly understood. Here, we developed deep learning models for dynamic recognition and compared different computational mechanisms, contrasting feedforward and recurrent, single-image and sequential processing as well as different forms of adaptation. We found that only models that integrate images sequentially via lateral recurrence mirrored human performance (N = 36) and were predictive of trial-by-trial responses across image durations (13-80 ms/image). Importantly, models with sequential lateral-recurrent integration also captured how human performance changes as a function of image presentation durations, with models processing images for a few time steps capturing human object recognition at shorter presentation durations and models processing images for more time steps capturing human object recognition at longer presentation durations. Furthermore, augmenting such a recurrent model with adaptation markedly improved dynamic recognition performance and accelerated its representational dynamics, thereby predicting human trial-by-trial responses using fewer processing resources. Together, these findings provide new insights into the mechanisms rendering object recognition so fast and effective in a dynamic visual world.

Manifold Learning for fMRI time-varying FC

Whole-brain functional connectivity ( FC ) measured with functional MRI (fMRI) evolve over time in meaningful ways at temporal scales going from years (e.g., development) to seconds (e.g., within-scan time-varying FC ( tvFC )). Yet, our ability to explore tvFC is severely constrained by its large dimensionality (several thousands). To overcome this difficulty, researchers seek to generate low dimensional representations (e.g., 2D and 3D scatter plots) expected to retain its most informative aspects (e.g., relationships to behavior, disease progression). Limited prior empirical work suggests that manifold learning techniques ( MLTs )-namely those seeking to infer a low dimensional non-linear surface (i.e., the manifold) where most of the data lies-are good candidates for accomplishing this task. Here we explore this possibility in detail. First, we discuss why one should expect tv FC data to lie on a low dimensional manifold. Second, we estimate what is the intrinsic dimension (i.e., minimum number of latent dimensions; ID ) of tvFC data manifolds. Third, we describe the inner workings of three state-of-the-art MLTs : Laplacian Eigenmaps ( LE ), T-distributed Stochastic Neighbor Embedding ( T-SNE ), and Uniform Manifold Approximation and Projection ( UMAP ). For each method, we empirically evaluate its ability to generate neuro-biologically meaningful representations of tvFC data, as well as their robustness against hyper-parameter selection. Our results show that tvFC data has an ID that ranges between 4 and 26, and that ID varies significantly between rest and task states. We also show how all three methods can effectively capture subject identity and task being performed: UMAP and T-SNE can capture these two levels of detail concurrently, but L E could only capture one at a time. We observed substantial variability in embedding quality across MLTs , and within- MLT as a function of hyper-parameter selection. To help alleviate this issue, we provide heuristics that can inform future studies. Finally, we also demonstrate the importance of feature normalization when combining data across subjects and the role that temporal autocorrelation plays in the application of MLTs to tvFC data. Overall, we conclude that while MLTs can be useful to generate summary views of labeled tvFC data, their application to unlabeled data such as resting-state remains challenging.