From task structures to world models: what do LLMs know?

Trends in Cognitive Sciences, Год журнала: 2024, Номер 28(5), С. 404 - 415

Опубликована: Март 4, 2024

Язык: Английский

---

Neurobehavioral meaning of pupil size

Neuron, Год журнала: 2024, Номер 112(20), С. 3381 - 3395

Опубликована: Июнь 25, 2024

Pupil size is a widely used metric of brain state. It one the few signals originating from that can be readily monitored with low-cost devices in basic science, clinical, and home settings. is, therefore, important to investigate generate well-defined theories related specific interpretations this metric. What exactly does it tell us about brain? Pupils constrict response light dilate during darkness, but also controls pupil irrespective luminosity. fluctuations resulting ongoing "brain states" are as arousal, what pupil-linked arousal how should interpreted neural, cognitive, computational terms? Here, we discuss some recent findings these issues. We identify open questions propose answer them through combination tasks, neurocomputational models, neurophysiological probing interconnected loops causes consequences size.

Язык: Английский

---

Better models of human high-level visual cortex emerge from natural language supervision with a large and diverse dataset

Nature Machine Intelligence, Год журнала: 2023, Номер 5(12), С. 1415 - 1426

Опубликована: Ноя. 13, 2023

Язык: Английский

---

A goal-centric outlook on learning

Trends in Cognitive Sciences, Год журнала: 2023, Номер 27(12), С. 1150 - 1164

Опубликована: Сен. 9, 2023

Язык: Английский

---

Rationality, preferences, and emotions with biological constraints: it all starts from our senses

Trends in Cognitive Sciences, Год журнала: 2024, Номер 28(3), С. 264 - 277

Опубликована: Фев. 9, 2024

Is the role of our sensory systems to represent physical world as accurately possible? If so, are preferences and emotions, often deemed irrational, decoupled from these 'ground-truth' experiences? We show why answer both questions is 'no'. Brain function metabolically costly, brain loses some fraction information that it encodes transmits. Therefore, if brains maximize objective functions increase fitness their species, they should adapt objective-maximizing rules environment at earliest stages processing. Consequently, observed 'irrationalities', preferences, emotions stem necessity for early process while considering metabolic costs internal states organism.

Язык: Английский

---

Jointly efficient encoding and decoding in neural populations

PLoS Computational Biology, Год журнала: 2024, Номер 20(7), С. e1012240 - e1012240

Опубликована: Июль 10, 2024

The efficient coding approach proposes that neural systems represent as much sensory information biological constraints allow. It aims at formalizing encoding a constrained optimal process. A different approach, decoding, instantiate generative model of the world. Here, we put forth normative framework characterizes jointly optimizing and decoding. takes form variational autoencoder: stimuli are encoded in noisy activity neurons to be interpreted by flexible decoder; must allow for an accurate stimulus reconstruction from activity. Jointly, is required statistics latent features which mapped decoder into distributions over stimuli; decoding correspondingly optimizes accuracy model. This yields family encoding-decoding models, result equally indexed measure stimulus-induced deviation marginal distribution Each member this predicts specific relation between properties neurons—such arrangement tuning curve means (preferred stimuli) widths (degrees selectivity) population—as function Our thus generalizes approach. Notably, here, constraint on optimization derives requirement model, while it arbitrary models. Moreover, solutions do not require knowledge distribution, but learned basis data samples; further acts regularizer, allowing generalize beyond training data. Finally, characterize models obtain through alternate measures performance, such error reconstruction. We find range admits comparable performance; particular, population with broad curves observed experimentally both low robustly unseen

Язык: Английский

---

An active inference perspective for the amygdala complex

Trends in Cognitive Sciences, Год журнала: 2023, Номер 28(3), С. 223 - 236

Опубликована: Дек. 15, 2023

We outline how a predictive processing framework based on active inference can help overcome the limitations of fragmented feed-forward theories and more comprehensively explain role amygdala in anxiety, fear, danger detection.This integrates theoretical predictions with empirical findings, suggesting that central subnucleus acts as Bayesian regulator an interoceptive self-model, sending top-down to basolateral for efficient perception.Fear conditioning is discussed example proactive homeostatic regulation, where exteroceptive cues are used anticipate negative experiences.By extension, other functions (e.g., its approach/avoidance behavior or attention) be explained using same computational principles within hierarchical model. The heterogeneous network subcortical nuclei importance cognitive clinical neuroscience. Various experimental designs human psychology animal model research have mapped multiple conceptual frameworks valence/salience decision making) ever refined circuitry. However, these predominantly bottom up-driven accounts often rely interpretations tailored specific phenomenon, thus preventing comprehensive integrative theories. argue here function could unify fractionated approaches into overarching clearer mechanistic interpretations. This embeds models, informed by prior knowledge belief updating, dynamical system distributed across circuits which self-regulation implemented continuously tracking environmental demands. subnuclei quintessential social, cognitive, affective, neuroscience [1.Adolphs R. What does contribute social cognition?.Ann. N. Y. Acad. Sci. 2010; 1191: 42-61Crossref PubMed Scopus (613) Google Scholar, 2.Bickart K.C. et al.The hub brain networks support life.Neuropsychologia. 2014; 63: 235-248Crossref (246) 3.Rosenberger L.A. indispensable experiential learning.Curr. Biol. 2019; 29: 3532-3537Abstract Full Text PDF (22) 4.Terburg D. essential rapid escape: rodent study.Cell. 2018; 175: 723-735Abstract (89) Scholar]. With neuroimaging circuit charting functional heterogeneity [5.Sladky al.Unsmoothed MRI bed nucleus stria terminalis during emotional faces.Neuroimage. 168: 383-391Crossref (21) 6.Sladky al.Basolateral orchestrate we learn whom trust.Commun. 2021; 4: 1-9Crossref (3) 7.Torrisi S. al.Extended connectivity changes sustained shock anticipation.Transl. Psychiatry. 8: 1-12Crossref (0) 8.Tillman R.M. al.Intrinsic extended amygdala.Hum. Brain Mapp. 39: 1291-1312Crossref (51) Scholar], our understanding has become detailed, yet fractionated. Moreover, some key such fear conditioning, do not easily replicate paradigms organisms [9.Visser al.Robust BOLD responses faces but conditioned threat: challenging amygdala's reputation extinction learning.J. Neurosci. 41: 10278-10292Crossref (15) problem lies data absence theory would allow unifying less situational ad hoc interpretation function(s). comprises several smaller [10.Janak P.H. Tye K.M. From behaviour amygdala.Nature. 2015; 517: 284-292Crossref (1205) (BLA) considered mainly stimulus sensory further divided two parts: lateral (LA) sensitive unimodal sensations [11.Uwano T. al.Neuronal responsiveness various stimuli, associative learning rat amygdala.Neuroscience. 1995; 68: 339-361Abstract (143) Scholar] connections cortices thalamus, whereas basal (BA) encodes multimodal contextual features [12.Yaniv al.A gradient plasticity revealed cortical stimulation, vivo.Neuroscience. 2001; 106: 613-620Abstract (32) behavioral [13.Kyriazi P. al.Multi-dimensional coding neurons.Neuron. 99: 1315-1328Abstract (64) hippocampus entorhinal cortex [14.Davis Reijmers L.G. dynamic nature engrams amygdala.Brain Res. Bull. 141: 44-49Crossref (18) allowing abstract encodings [15.Duvarci Pare Amygdala microcircuits controlling learned fear.Neuron. 82: 966-980Abstract (523) (CeA) reflect general motivational aspects [16.Balleine B.W. Killcross Parallel incentive processing: integrated view function.Trends 2006; 272-279Abstract (452) Scholar,17.Shackman A.J. al.Dispositional negativity: psychological neurobiological perspective.Psychol. 2016; 142: 1275-1314Crossref (109) and, together distal (BST), might encode features, threat (Table 1, Key table). Although cell populations marked differences types, neurotransmitter binding sites, gene expression, [18.Babaev O. al.Inhibition anxiety circuitry.Exp. Mol. Med. 50: 1-16Crossref (168) Scholar,19.Hur J. al.Anxiety neurobiology temporally uncertain anticipation.J. 2020; 40: 7949-7964Crossref (48) inhabit small space, differentiation found noninvasive fMRI, between BLA CeA [6.Sladky BST [7.Torrisi given high fMRI sensitivity ultra-high field fMRI) and/or well-powered tasks providing sufficient specificity.Table 1Key table. Feed-forward versus selected topicsAnatomical diagramFeed-forward processingActive inferenceLA connected well BA. connects direct indirect pathways (via insula forebrain)Theory integrate stimuli gate behaviorCeA self-model passes valence BLA, tune predictionsTopic 1: detection BLASensory signaled CeA. How filter relevant features?Sensations actively filtered (suppressed enhanced) thalamus areas, sensingTopic 2: CeASensations trigger feature makes particular sensation fear-inducing?CeA form self-regulatory process. Fear deviations homeo-/allostatic set-pointsTopic 3: conditioningNeutral associated feared stimulus. affective value assigned previously neutral stimuli?(Harmless) predict from priors adaptive behaviorTopic approach avoidance causes somatic changes. actions gated regulated?Actions parallel processes levels (BLA, attention; CeA, stereotypical behavior; cortex, flexible behavior) Open table new tab Commonly, presented following principles: after operates relay output region motivated [20.Davis M. Whalen P.J. amygdala: vigilance emotion.Mol. 6: 13-34Crossref (2347) Scholar,21.LeDoux J.E. Emotion brain.Annu. Rev. 2000; 23: 155-184Crossref (6385) Modern views appreciate complexity executive control information integration [22.Fadok J.P. al.New perspectives function.Curr. Opin. Neurobiol. 49: 141-147Crossref (148) still face substantial limitations. Foremost, at what stage mere transformed meaning organism danger, safety, reward)? different topics 1), typically studied isolation, unified under one biological/computational principle? Empirical aimed relating microcircuit dynamics basic primitives affectivity positive/negative valence; see Glossary) active/passive responding) Scholar,23.Fadok competitive inhibitory selection passive responses.Nature. 2017; 542: 96-100Crossref (267) 24.Gozzi A. neural switch 67: 656-666Abstract (150) 25.Pliota al.Stress peptides sensitize circuitry promote coping.Mol. 25: 428-441Crossref (11) limited success attribute traditional perspective. First, even carefully planned experiments, humans animals exhibit striking variability their behavior, pointing toward considerable intersubjective state dependencies responses. handled treating uninformative noise, essentially means discarding part observed data; alternatively, additional confounding variables obscure apparent neutrality behavioristic research. Then, no longer describes naturally observable relationship response: it contaminated artifactual assumptions. Second, mounting evidence resting-state reveals intrinsic activity independent external perturbations condition fundamental [26.Coste C.P. al.Ongoing fluctuations directly account intertrial indirectly intersubject Stroop task performance.Cereb. Cortex. 2011; 21: 2612-2619Crossref Scholar,27.Sadaghiani large-scale perception.Proc. Natl. U. 112: 8463-8468Crossref (163) Third, level subjective, first-person experience, assumption cognition only triggered violates validity. All necessary corollary steps imply that, sake operationalization, important usually accounted for, neither nor practice. Stimulus-driven, models replaced action-oriented self-regulation, driven higher hierarchy states lower (Box 1). If input diverges predictions, prediction error signal passed back level. In this way, novel surprising relayed [28.Clark Whatever next? Predictive brains, situated agents, future science.Behav. 2013; 36: 181-204Crossref (2893) resulting computationally [29.Ali al.Predictive consequence energy efficiency recurrent networks.Patterns. 2022; 3100639Abstract (8) Scholar,30.Rao R.P.N. Ballard D.H. visual cortex: extra-classical receptive-field effects.Nat. 1999; 79-87Crossref (3074) [31.Friston K.J. Stephan K.E. Free-energy brain.Synthese. 2007; 159: 417-458Crossref (409) 32.Friston K. free-energy principle: rough guide brain?.Trends Cogn. 2009; 13: 293-301Abstract (1084) 33.Parr al.Active Inference: Free Energy Principle Mind, Brain, Behavior. MIT Press, 2022Crossref regulation 2). Prediction errors known reinforcement rewards [34.Hollerman J.R. Schultz W. Dopamine neurons report temporal reward learning.Nat. 1998; 304-309Crossref Scholar,35.Mirenowicz Preferential activation midbrain dopamine appetitive rather than aversive stimuli.Nature. 1996; 379: 449-451Crossref (661) generalized perception action [32.Friston Scholar,36.Friston epistemic value.Cogn. 187-214Crossref (432) emotions applying concepts self-models [37.Seth A.K. Friston Active brain.Philos. Trans. Soc. B. 37120160007Crossref (448) constructed [38.Barrett L.F. emotion: interoception categorization.Soc. Affect. 12: 1-23Crossref (142) variational free [39.Joffily Coricelli G. Emotional principle.PLoS Comput. 9e1003094Crossref (171) 3). Expanding ideas, propose inference, illustrate four 1).Box 1Free energy, processing, brainFoundational principle assumes all nervous follow minimizing (variational) Fμαy, internal μ, α, observations y. To infer statistical regularities world, agent creates generative (m; i.e., thoughts, fantasies), tested against available (y; input, data, sensations, evidence). neurocomputational implementation hierarchical, is, send down hierarchy, while up (predictive processing). continuous bidirectional process refines reduces discrepancies expectations reality (prediction minimization). Higher slower environment (temporal depth). Going beyond passively updating coding), also act upon world make conform (i.e., change inputs receives interacting environment).Minimizing minimizes surprise Fμαy≥−lnpym,α, reduction, making special formulation processing. inverse evidence, optimization comparison, tractable biologically plausible method implementing building blocks brain. Given term reformulated Fμαy=Complexity–Accuracy minimized ways. creating sparser, efficient, generalizable Complexity. Accuracy related achieved developing accurate (m) reliably predicts (y). Actions (α) affect accuracy ways: (i) sampling way modulation attention moving eyes away stimulus) sensed selectively enhancing suppressing distinguish noise; (ii) come true prefers going dark forest turns flashlight). themselves constantly guided goals perceptual (to improve predictions) instrumental verify predictable sensation) [68.Seth Tsakiris Being beast machine: basis selfhood.Trends 22: 969-981Abstract (see Box 2 main text).Box 2Active self-regulationActive described 1 text, brain, understood finding approximate optimal solution (integrating beliefs) choice policy beliefs. Computationally, precision-weighted [49.Mathys C.D. al.Uncertainty Gaussian filter. Front. Hum, Neurosci2014: 8Google posits weight depends ratio precision (or variance, π=σ−2) [e.g., px∼Nμxπx−1] likelihood pyx∼Nμyπy−1]. posterior distribution pxy∼Nμx∣yπx∣y−1 then calculated πx∣y=πx+πy mean μx∣y=μx+πyπx∣yy−μx [i.e., = + rate) × error] Figure text).On level, weighting thought neuromodulatory mechanisms example, acetylcholine tuning synaptic gain consequently, adjust rate. Updating called when occurs short timescales single observation), parameter updates trials). For perception, rate should highest if low situations) originates precise seeing object very clearly). lowest knowing one's living room) unreliable dark). distinct 'active' inference:First, sample features. Early eye-tracking studies had already showed participants images differently depending [112.Yarbus A.L. Eye Movements Vision. Springer, 1967Crossref famous basketball gorilla experiment [113.Simons D.J. Chabris C.F. Gorillas midst: inattentional blindness events.Perception. 28: 1059-1074Crossref focused leads attentional highly unusual events (low priori probability context). Even though useful they price introducing well-known biases, illusions exploit (seeing you seeing), confirmation bias want knowledge), overgeneralizations (not there). Our dependence why false beliefs about feel were real [114.Dijkstra Fleming S.M. Subjective strength distinguishes imagination.Nat. Commun. 2023; 14: 1627Crossref (7) implications biased disorders text).Second, regulation. high, parameters simply updated match case epistemically coding. These flexible, preference correspond metabolic, allostatic, autopoietic needs organism, body state, preferable being safe space vs. exposed predation). depend objective (how is), subjective preferences be)Box 3Emotions: descriptive naturalisticThe concept construction due success. perspective provides alternative action-/goal-/outcome-oriented embodied linked especially Scholar]). enables us reformulate affectivity.Salience results relevance, mediated weighting. Valence estimating current expected Formally, divergence proportional inspire testable, [103.Hesp C. al.Deeply felt affect: emergence deep inference.Neural 33: 398-446Crossref (71) been suggested (expected) increase disappointment, unhappiness, decrease relief, hope, happiness Altogether, emotions, reward, emergent consequences principles. Ideally, seeks out experiences predicted avoids those result feeling too cold hot, Consequently, (subjectively) rewarding preferably detected [115.Banerjee al.Value-guided remapping orbitofrontal cortex.Nature. 585: 245-250Crossref (70) Scholar,116.Schaffner al.Sensory relies fitness-maximizing codes.Nat. Hum. Behav. 7: 1135-1151Crossref automatic harm elicited extremely simple [117.De Franceschi al.Vision guides freeze flight defense strategies mice.Curr. 26: 2150-2154Abstract (156) preferred states, shifts focus function. There genetically defined modality population (Figure 3 [69.Kim behaviors.Neuron. 93: 1464-1479Abstract (256) Scholar]), encoding currently unclear.Thevalue drives [118.Sharpe M.J. al.Dopamine transients acquisition model-based associations.Nat. 20: 735-742Crossref reward-sensitive [89.Warlow recruits mesocorticolimbic pursuit pain.Nat. 11: 2716Crossref (31) pleasure. Instead, rodents engage optogenetic self-stimulation [100.Warlow Berridge Incentive motivation: 'wanting' roles circuitry.Behav. 411113376Crossref (29) require concurrent supports links inference.Arousal via neuromodulation, controlled slight hunger critical compared discomfort pain, death). implement asymmetric p

Язык: Английский

---

Approach-avoidance conflict recruits lateral frontoparietal and cinguloinsular networks in a predator-prey game setting

Cognitive Affective & Behavioral Neuroscience, Год журнала: 2025, Номер unknown

Опубликована: Фев. 26, 2025

Язык: Английский

---

Neural waves and computation in a neural net model I: Convolutional hierarchies

Journal of Computational Neuroscience, Год журнала: 2024, Номер 52(1), С. 39 - 71

Опубликована: Фев. 1, 2024

Язык: Английский

---

Endogenous Precision of the Number Sense

Опубликована: Сен. 20, 2024

The behavioral variability in psychophysical experiments and the stochasticity of sensory neurons have revealed inherent imprecision brain’s representations environmental variables 1–6 . Numerosity studies yield similar results, pointing to an imprecise ‘number sense’ brain 7–13 If reflects optimal allocation limited cognitive resources, as suggested by efficient-coding models 14–26 , then it should depend on context which are elicited 25,27 Through estimation task a discrimination task, both involving numerosities, we show that scale subjects’ increases, but sublinearly, with width prior distribution from numbers sampled. This sublinear relation is notably different two tasks. double dependence — consistent optimization tradeoff between expected reward, for each resource cost encoding neurons’ activity. Comparing tasks allows us clarify form constraint. Our results suggest perceptual noise endogenously determined, precision percepts varies they elicited, observer’s objective.

Язык: Английский

---