Neural activity unfolds in time, yet many of its governing principles appear strikingly symmetric when viewed from the perspective of dynamical systems rather than moment-to-moment causes and effects. Time symmetry in neural dynamics refers to the idea that the rules describing how neural states evolve can often be written in forms that are invariant, or nearly invariant, under reversal of the time axis. This does not mean that individual spikes literally run backward, but that the same underlying generative laws can account for how activity patterns propagate both from past to future and, in certain formal senses, from future constraints back to earlier states. Such symmetry becomes apparent when describing neural computation in terms of trajectories through high-dimensional state spaces, where constraints imposed at both initial and final times help determine the entire path in between.
Energy-based formulations of neural networks provide a natural language for expressing these ideas. In these formulations, patterns of neural activity correspond to points in an energy landscape, and the dynamics consist of movements downhill toward lower-energy configurations that satisfy more constraints. When the update rules are derived from gradients of a scalar energy function, the equations governing the evolution of the system can be written so that they do not explicitly privilege a forward or backward direction in time; what matters instead is how the system moves between attractor states that locally minimize energy. This perspective suggests a kind of effective time symmetry: the same energy function that predicts how a network relaxes from an initial perturbation can also describe how final constraints, such as a target pattern of activity, shape the permissible history of states leading up to it.
From a statistical perspective, time-symmetric principles emerge naturally when neural dynamics are interpreted as a physical realization of bayesian inference. In this view, latent causes in the world give rise to sensory data, and the brain’s task is to infer those causes by combining priors with incoming evidence. The generative model that links causes to observations is typically formulated without privileging a temporal direction: it specifies joint probabilities over sequences of states and data. Bayes’ rule then links forward-time generative processes with backward-time inference processes through a symmetric mathematical structure. Although the physical influences on sensory receptors come from the past, the internal computations that decode those influences can be described as bidirectional message passing, where information from future observations can retroactively reweight interpretations of earlier events within a given observation window.
Continuous-time descriptions of cortical population activity, such as those based on stochastic differential equations, further illuminate how time-symmetric principles appear in neural dynamics. In these models, the evolution of firing rates or membrane potentials is characterized by drift and diffusion terms that can often be related to a potential function. When such a potential exists, the forward-time Fokker–Planck equation describing the probability flow over states has a corresponding backward equation that determines how boundary conditions at later times influence expectations at earlier times. This duality captures a formal time symmetry: the same underlying dynamics support both forward propagation of uncertainty and backward propagation of constraints, and neural trajectories can be understood as reconciling these two directions of influence.
Spike-based formulations, though inherently discrete and seemingly arrow-of-time–laden, can also be recast in time-symmetric terms when analyzed as point processes embedded in a larger probabilistic structure. A spike train can be seen as one realization of a stochastic process governed by an intensity function that depends on both past and, in a conditional sense, future context within a trial. When fitting such models, researchers frequently use smoothing methods that estimate latent firing intensities based on both earlier and later spikes, effectively treating the spike train as a whole and using temporally symmetric inference procedures. The resulting estimates obey dynamical laws that are symmetric under time reversal at the level of the latent variables, even though the spikes themselves only occur in the forward direction.
Recurrent neural networks offer another setting in which time-symmetric ideas become explicit. In their simplest deterministic form, the next state of the network is a function of the current state and input. However, when these networks are trained to represent probability distributions or to implement evidence integration over time, the learned weights often encode constraints that can be satisfied by running the same network dynamics forward or backward along a sequence. Autoencoders that process temporal data, for instance, can be constructed so that the encoding and decoding phases are related by approximate inverses, mirroring a symmetry between compressing past information and reconstructing it from future constraints. This near-reversibility implies that the essential computational principles governing state transitions do not depend strongly on the direction of traversal.
In the context of predictive processing, time-symmetric principles appear in how the brain balances top-down predictions with bottom-up prediction errors. At any moment, cortical circuits are thought to maintain hypotheses about the causes of sensory inputs, then adjust those hypotheses when mismatches arise. Formally, this process resembles iterative optimization of a cost function defined over entire temporal sequences, where predictions and priors interact with incoming data. The gradient-based rules that adjust neural activity or synaptic strengths can be derived from objectives that treat past and future observations within a trial on an almost equal footing. In such formulations, the updates that would be computed if one had access to all future data can be expressed in a mathematically similar way to those based only on past data, reinforcing the notion that the underlying inference principles themselves are time-symmetric, even if the brain approximates them in a causal, online fashion.
Time-symmetric neural dynamics also appear in the relationship between encoding and decoding of information. Neural codes are often analyzed by asking how well one can reconstruct a stimulus from recorded activity, or conversely, how a given stimulus drives the network. When the mapping between stimuli and responses is modeled as part of a joint probability distribution, the encoding and decoding problems become duals of each other, linked by the same underlying structure. This duality is inherently symmetric: swapping the roles of causes and effects corresponds to traversing the same statistical relationships in reverse. Neural systems that approximate such optimal encoding–decoding schemes can thus be described as implementing computations that are symmetric with respect to time, in the sense that the same internal model supports both forward generation of responses and backward inference of causes.
At the microscopic level, biophysical processes governing ion channels and synapses are rooted in fundamental physical laws that are themselves time-reversal symmetric, at least to a very good approximation in the regimes relevant to biology. While thermodynamic irreversibility and metabolic costs impose a macroscopic arrow of time on neural tissue, the local interactions between molecules respect equations that do not change form when time is reversed. This physical substrate allows, in principle, for neural circuits to implement algorithms derived from symmetric variational principles—such as minimizing a free energy functional defined over trajectories—where optimal paths can be characterized without specifying a preferred temporal direction. The observed directionality in perception and action then emerges not from the basic dynamical rules but from boundary conditions, such as sensory inputs and behavioral goals.
These observations together motivate viewing neural computation as unfolding on a space of possible histories, constrained at both ends by what has already occurred and what must eventually be satisfied, such as task demands or homeostatic set points. In this trajectory-based viewpoint, time-symmetric principles are embodied in optimization problems defined over entire paths rather than isolated moments. Neural states at intermediate times are selected to jointly respect constraints propagating from past conditions and from anticipated or imposed final states. The concrete, causal operation of neurons and synapses thus realizes, in an approximate and metabolically constrained manner, algorithms that are most naturally expressed in a time-symmetric mathematical language, bridging the apparent gap between irreversible biological processes and symmetric computational formalisms.
Bidirectional information flow in cortical circuits
Cortical circuits are anatomically and functionally organized to support information flow in both feedforward and feedback directions, and this bidirectionality is central to how neural computation expresses time symmetry at the circuit level. Sensory areas receive ascending inputs that convey rapidly changing features of the environment, while descending pathways from higher-order regions carry contextual signals, expectations, and task demands. Rather than acting as separate channels, these streams of activity converge on shared populations of neurons, whose membrane potentials and firing patterns simultaneously reflect bottom-up drive and top-down constraints. In this arrangement, the state of a cortical column at any given moment encodes a compromise between evidence arriving from the past via the sensory periphery and constraints that effectively point toward anticipated future states of the system.
The laminated structure of cortex makes this functional reciprocity explicit. Thalamic and lower-level cortical inputs predominantly target middle layers, while feedback from higher association areas and frontal regions preferentially innervates superficial and deep layers. Neurons in these layers, in turn, project both upward and downward in the hierarchy, forming recurrent loops that can sustain and refine activity patterns over extended time windows. This architecture implements a form of evidence integration in which signals originating at different hierarchical levels, and therefore associated with different temporal horizons, are continually blended. Ascending spikes bring in information about what has just occurred, whereas descending spikes encode predictions and priors about what should be happening, allowing the same synaptic machinery to reconcile information that is effectively oriented toward both past causes and future consequences.
Local microcircuits further reinforce this bidirectional organization through specific patterns of excitation and inhibition. Pyramidal neurons projecting to higher-order regions receive feedback from those very targets, closing trans-cortical loops that allow activity patterns to reverberate between areas. Interneurons with distinct molecular and electrophysiological profiles selectively gate these interactions, shaping when and how top-down signals can influence ongoing processing. For example, inhibitory cells that target distal dendrites can modulate the impact of feedback projections impinging on those dendritic compartments, while other interneuron types regulate the timing and gain of feedforward drive arriving at perisomatic regions. The net effect is that cortical responses at any instant are the product of a dynamically reconfigured balance between inputs conveying immediate sensory evidence and inputs shaping long-range contextual alignment, rather than a simple forward cascade.
From a functional standpoint, feedforward paths can be viewed as carrying data likelihood information, while feedback paths convey expectations derived from internal models. Within this bayesian inference perspective, cortical circuits approximate message passing algorithms, where evidence moves up the hierarchy and predictions move down. The same synapses that transmit bottom-up feature activations also participate in transmitting top-down error corrections or hypothesis confirmations, depending on the current state of the network. This dual use of circuitry suggests that the underlying rules governing synaptic integration are not intrinsically tied to one temporal direction; instead, they specify how local membrane potentials should be updated given a mixture of messages, irrespective of whether those messages are interpreted as coming “from the past” or “from the future” of the representational hierarchy.
Bidirectional processing also manifests in the laminar flow of prediction errors and predictions envisioned in predictive coding schemes. In these models, distinct neuronal populations within each cortical area encode the current best estimate of latent causes and the discrepancy between that estimate and incoming signals. Estimated causes send predictions downward, while prediction error units send residuals upward. Biophysically, both types of signals are carried by action potentials using the same ion channels and synaptic mechanisms; the difference lies in the circuit motifs and the computational roles assigned to particular projection patterns. Because the same generative model underlies both the downward propagation of predictions and the upward propagation of errors, the mathematical structure connecting causes, observations, and internal states is symmetric under an exchange of roles, even though physical propagation remains forward in time.
Top-down influences are not purely modulatory; they can dramatically reshape the trajectory of activity in early sensory cortices, effectively reinterpreting the same feedforward input under different contextual assumptions. Attention, expectation, and task rules all exert their effects by adjusting the gain and selectivity of early-stage neurons through feedback. For instance, anticipatory activity in visual cortex can preconfigure receptive fields before a stimulus appears, biasing subsequent responses toward interpretations consistent with prior knowledge. In this sense, feedback conveys information about constraints that are anchored in anticipated or desired future states of the organism, such as upcoming actions or rewards, and these constraints propagate backward through the hierarchy to inform how current sensory data should be parsed. The resulting neural trajectories, measured over short time scales, embody a compromise between what has already been sensed and what must be satisfied later, as encoded in goals and learned regularities.
Even within a single cortical area, recurrent connections mediate bidirectional flow between neuronal subpopulations that process temporally distinct aspects of a stimulus. Cells with short-latency responses may quickly encode coarse, high-contrast features based on the first waves of afferent spikes, while slower, more integrative populations refine those representations as additional evidence accumulates and feedback arrives. Recurrent loops among these populations enable information initially bound to early segments of a stimulus to be revisited and updated in light of patterns emerging later in the trial. This continual re-entrance of earlier states into ongoing computation embodies a form of circuit-level time symmetry: the network does not simply march forward through a fixed sequence of states but repeatedly revises and partially reuses prior activity patterns to accommodate new constraints.
Evidence for bidirectional information flow is especially clear in tasks requiring the reconciliation of ambiguous or noisy inputs with stable expectations. In such settings, cortical activity often exhibits late components that are more strongly tied to the eventual perceptual report than to the raw stimulus features. These late signals, propagating down from decision-related or associative regions, reshape the activity in sensory cortices to align it with the chosen interpretation, effectively “back-projecting” the outcome of higher-level inference onto earlier processing stages. Although no literal retrocausality occurs, the pattern of activation in lower areas at a given time can reflect constraints that originate in processes which, from a behavioral perspective, are closer to future decisions than to past sensory events. Circuit dynamics thus encode information that is jointly constrained by where the system came from and where it is headed.
Bidirectional pathways also support flexible routing of information across cortical areas, enabling different subnetworks to transiently couple or decouple depending on task demands. During working memory and decision-making, for example, prefrontal regions can selectively amplify or suppress specific sensory representations through feedback, effectively steering which aspects of the past input will continue to influence ongoing processing and which will be dampened. Conversely, strong feedforward bursts from unexpected stimuli can override ongoing top-down patterns, forcing a rapid revision of predictions and priors. The same structural connections participate in both regimes: at times dominated by top-down control and at times dominated by bottom-up drive. This flexibility indicates that the computational roles of particular pathways are contingent on the current state of the network and the temporal context, rather than being permanently tied to one direction of information flow.
Oscillatory synchronization provides another mechanism for coordinating bidirectional exchange between cortical regions. Rhythms in different frequency bands have been associated with preferential feedforward or feedback signaling, but these associations are not rigid. Phase relationships between oscillations can dynamically gate when spikes from one area most effectively drive neurons in another, producing windows of enhanced influence that can alternate between ascending and descending dominance. As a result, over a full cycle of coordinated rhythms, information may be pushed up the hierarchy during certain phases and pulled down during others, yielding an effectively bidirectional communication protocol that unfolds on the timescale of tens to hundreds of milliseconds. Neural computation in such regimes is best understood in terms of cycles of mutual constraint-satisfaction rather than strictly sequential stages.
At the level of synaptic plasticity, bidirectional information flow shapes how connections encode experience. Synapses on dendritic segments that primarily receive feedback become tuned to patterns that predict the future states of their postsynaptic targets, while synapses on segments dominated by feedforward input encode patterns tied more directly to recent sensory history. Because both sets of synapses converge onto the same neurons, learning rules must reconcile predictions carried by feedback with discrepancies revealed by feedforward drive. Hebbian mechanisms, spike-timing–dependent plasticity, and neuromodulatory signals such as dopamine collectively adjust these weights so that over time, the circuit’s dynamics come to encode regularities that link past inputs to future outcomes. The resulting connectivity patterns support computations in which present activity simultaneously serves as an explanation of what just happened and as a preparation for what is likely to occur next.
Bidirectional information flow in cortical circuits, therefore, is not an incidental byproduct of wiring complexity but a fundamental organizational principle that allows neural activity to be shaped by constraints propagating from both temporal directions. Feedforward and feedback pathways jointly sculpt a high-dimensional state space in which each trajectory is constrained by sensory evidence inherited from the past and by goals, expectations, and task structures aligned with the future. In this view, time symmetry in cortical processing arises at the algorithmic level: the same anatomical loops and synaptic mechanisms support both the upward propagation of data and the downward imposition of structure, allowing neural computation to approximate inference over entire temporal sequences rather than processing each moment in isolation.
Computational models of temporally symmetric inference
Computational models that explicitly embody temporally symmetric inference typically start from a generative description of how latent states and observations coevolve over time. In such models, the world and the brain are described jointly by probability distributions over entire trajectories, rather than by stepwise updates that only depend on the immediate past. The key move is to treat a sequence of neural states and sensory inputs as a single object governed by a time-symmetric law, then derive both forward predictions and backward inferences from this shared structure. This contrasts with traditional causal models, which privilege the forward direction and then bolt on separate mechanisms to account for retrospective interpretation. When the generative model itself is defined over complete paths, bayesian inference naturally becomes a temporally extended process in which observations at later times can influence beliefs about earlier hidden states without invoking any physical retrocausality.
A canonical example arises in state-space models used to describe population activity and behavior. Here, latent variables represent underlying causes such as stimulus features, task states, or internal goals, while observed variables correspond to spike trains, local field potentials, or behavioral outputs. The joint distribution over latent and observed sequences is often specified through Markovian dynamics, where each latent state depends on its neighbors in time and generates observations through a likelihood function. Importantly, this joint distribution is symmetric in the sense that it does not encode a preferred direction of explanation: it assigns probabilities to full sequences, regardless of whether one thinks of them as generated forward or reconstructed backward. Algorithms such as Rauch–Tung–Striebel smoothing or forward–backward message passing compute posterior distributions over latent trajectories by combining information from both temporal directions, embodying evidence integration that is mathematically time-symmetric even if implemented in a sequential machine.
Graphical models make this symmetry particularly transparent. When temporal processes are drawn as chains of nodes connected by probabilistic dependencies, the same factorization that defines forward dynamics also supports backward inference. Messages propagate along the chain from past to future and from future to past, updating beliefs at each intermediate time point. These messages correspond to likelihood terms contributed by observations and prior terms contributed by dynamic constraints, and their combination yields a posterior that is consistent across the entire sequence. In models of neural computation that adopt this framework, activities in a given time bin are interpreted as encoding beliefs that have already been shaped by both earlier and later evidence within the trial. The model itself is indifferent to whether information is said to “flow” forward or backward; that distinction only appears when mapping the computation onto physical circuitry with causal constraints.
Energy-based and Hopfield-type networks extend these ideas to spatially distributed patterns of activity, and they can be generalized to temporal inference by introducing time-indexed copies of the network. In such constructions, each time slice is represented as a layer of units, with symmetric connections not only within a slice but also between neighboring slices. A global energy function is then defined over the entire spatiotemporal configuration, and the goal of inference is to find configurations that minimize this energy subject to observed constraints at particular times. Because the couplings are symmetric, the gradient of the energy with respect to a given unit depends on its neighbors both in space and time in the same way, regardless of whether those neighbors are in the past or future index. Optimization methods such as contrastive divergence or equilibrium propagation can be applied to this extended network, yielding updates that are formally invariant under reversal of the time dimension as long as boundary conditions are transformed accordingly.
Equilibrium propagation and related frameworks have been proposed as biologically plausible approximations to backpropagation in recurrent networks. In these models, a network is first allowed to relax to a free equilibrium under a given input, then gently nudged toward a target output by applying a small perturbation. The difference between the free and nudged states encodes gradient information that can be used to update synaptic weights. When extended to temporal settings, the perturbation can be anchored at a later time point—such as a desired decision or reward—and allowed to diffuse backward through recurrent dynamics, modifying neural activity at earlier times before learning is consolidated. The same dynamical rules govern both the forward relaxation under input and the backward relaxation under constraint, revealing an implicit time symmetry in the learning algorithm: target signals imposed at the end of a trial shape internal states throughout the preceding interval via the same physics that previously shaped those states from initial conditions.
Probabilistic population codes provide another route to temporally symmetric inference. In these models, patterns of firing rates across a population represent probability distributions over latent variables, and neural operations such as linear pooling and divisive normalization implement approximate bayesian inference. When extended to time-varying stimuli, latent trajectories are inferred using both prior dynamics and incoming data. The prior relates neighboring time points, while the likelihood terms encode how each moment’s observations depend on the latent state. Filtering corresponds to using past data and priors to infer the current state, whereas smoothing incorporates future observations as well. In computational models that treat population responses as encoding smoothed posteriors, the effective neural code at any time reflects information drawn symmetrically from all observations in the window, blurring the distinction between encoding “what has just happened” and encoding “what will later be observed” during that same trial.
Variational inference offers a flexible framework in which time-symmetric principles are especially clear. Instead of computing exact posteriors over latent trajectories, one posits a family of approximate distributions and optimizes their parameters to minimize a divergence from the true posterior. When the variational family is structured as a bidirectional recurrent network, with separate forward and backward encoders that read the sequence in opposite directions, the resulting approximate posterior is explicitly constructed by combining information from both past and future time steps. This is the logic behind bidirectional recurrent neural networks and transformers with full-sequence attention, which use context from the entire sequence to represent each element. Although these architectures are typically trained for tasks such as sequence labeling or language modeling, they instantiate a general mechanism for temporally symmetric inference: internal representations at a given position are defined as functions of the whole trajectory, not just its past.
In models aimed at biological plausibility, the explicit use of bidirectional passes is often replaced by recurrent dynamics that gradually incorporate constraints over time. For instance, networks implementing predictive coding minimize a free energy or prediction error functional defined over an entire sequence, with separate populations encoding causes and errors. If the objective function is written as a sum of squared prediction errors across time, plus dynamic priors linking successive latent states, then gradient-based updates to latent activities involve error signals from both earlier and later observations. In principle, an idealized system performing exact gradient descent on this objective would exhibit time symmetry: the equations governing how a latent state is nudged by mismatches at adjacent times do not intrinsically favor one direction. Approximations that implement these updates in a causal manner, such as local message passing rules, can still reflect this underlying symmetry in their stationary solutions, even if the path taken to those solutions respects the arrow of physical time.
Path-integral and variational path inference methods push the symmetry perspective even further by treating entire neural trajectories as single objects over which probability mass is distributed. In such formulations, one defines an action functional that assigns a cost or log-probability to each possible trajectory, incorporating both dynamic smoothness terms and observation likelihoods. Optimal trajectories are those that minimize this action, and fluctuations around them are governed by the same functional. Because the action depends only on the trajectory as a whole, and not on a particular direction of traversal, the associated Euler–Lagrange or Hamiltonian equations are often invariant under time reversal up to boundary terms. Computational models that adopt this formalism view neural dynamics as solutions to a two-point boundary value problem, where both initial conditions (e.g., prior states or baseline activity) and terminal conditions (e.g., task success or reward delivery) constrain the intermediate path. The resulting solutions embody time symmetry at the level of the governing equations, even though real neural tissue can only explore them in a forward-time manner.
Reinforcement learning with eligibility traces illustrates how learning rules can be interpreted in temporally symmetric terms. In temporal-difference methods, prediction errors computed at later times propagate backward to update values associated with earlier states. Eligibility traces accumulate recent activity and maintain a decaying memory of which synapses or states were recently involved, while delayed rewards or errors determine the sign and magnitude of updates. When formalized probabilistically, this process corresponds to computing gradients of a return functional defined over full trajectories, which is intrinsically time-symmetric: the same objective describes how early choices influence late outcomes and how late outcomes should influence the evaluation of early choices. Biophysically motivated models that link eligibility traces to synaptic tags and neuromodulatory signals can therefore be seen as implementing a physically constrained approximation to a symmetric credit-assignment computation over time.
Computational models of working memory and decision-making often adopt attractor dynamics that are governed by symmetric connectivity matrices. These networks exhibit multiple stable or metastable states corresponding to different hypotheses, choices, or stored items. When such networks are embedded in a temporal context—for instance, during a delay-period task—their dynamics are shaped by constraints imposed both at stimulus onset and at response time. The network must evolve from an initial neutral state to a final decision state while satisfying constraints on stability, noise tolerance, and energy consumption. When described in terms of an effective potential landscape over trajectories, the same landscape determines which histories are plausible given a particular final attractor, regardless of whether one imagines the system rolling forward into the attractor or “unrolling” backward from it. Inference over which internal trajectory most likely produced an observed pattern of spikes therefore relies on the same dynamical laws that govern forward evolution, highlighting a form of time symmetry in the underlying computational model.
Sequence learning models, such as those based on hidden Markov models, recurrent networks, or synfire chains, provide additional examples. Many such models learn both forward and backward transition probabilities, or equivalently, learn a joint distribution over ordered pairs of states that can be queried in either direction. When used for decoding neural activity, backward transition probabilities allow one to infer likely preceding states from current observations. In statistical practice, smoothing algorithms that incorporate both forward and backward passes invariably outperform purely causal filtering when reconstructing latent neural dynamics from noisy data, reflecting the added information carried by future observations. From the viewpoint of time symmetry, these models show that the same learned structure—transition matrices, recurrent weights, or synaptic chains—simultaneously supports prediction of future activity and retrodiction of past activity, with no intrinsic asymmetry in the learned parameters themselves.
Transformers and attention-based models, though originally developed for artificial sequence processing, offer a powerful conceptual template for temporally symmetric neural computation. Self-attention mechanisms compute interactions between all pairs of time points in a sequence, allowing representations at each position to incorporate information from both earlier and later positions. When attention is not restricted to past tokens, the resulting model is explicitly noncausal and fully time-symmetric at the level of its internal computations. Each internal representation can be viewed as an optimal aggregator of evidence drawn from the entire temporal context, akin to a smoothed posterior in bayesian inference. Although such full-sequence attention is not directly compatible with strict online processing in biological systems, it clarifies how a unified model can, in principle, treat all time points on an equal footing and then impose causal constraints only at the stage of interaction with the external world.
Hybrid models that combine generative and discriminative components illustrate how time-symmetric principles can coexist with task-driven optimization. For example, a generative model may be trained to capture joint distributions over neural activity and stimuli across time, while a separate readout network is trained to perform classification or control using representations drawn from the generative model. The generative component maintains time symmetry at the level of its internal inference—using both past and future observations to shape latent trajectories—while the readout operates causally, relying only on information available up to the current time in a real task. This separation of concerns allows the system to leverage the statistical benefits of temporally symmetric inference during offline learning or replay, then deploy approximated, forward-only policies during online behavior. In such architectures, time symmetry resides in the inference and learning machinery, even as the outward expression of behavior respects the unidirectional flow of physical time.
Experimental evidence for time-symmetric neural processing
Empirical support for time-symmetric neural processing has emerged from multiple experimental paradigms that probe how brain activity reflects both earlier sensory inputs and later behavioral outcomes within a single trial. One of the clearest signatures comes from analyses showing that neural responses at intermediate times carry information not only about what has already been presented but also about what the subject will eventually perceive or decide. In perceptual decision-making tasks, for example, population activity in sensory and association cortices can be decoded to reveal the upcoming choice well before it is reported, and in some cases even before the critical sensory evidence has fully unfolded. This anticipatory tuning suggests that ongoing neural computation is constrained simultaneously by incoming inputs and by task-specific endpoints, consistent with a framework in which internal states are shaped by boundary conditions distributed across the temporal extent of a trial rather than by purely feedforward drive.
Reverse-correlation and choice-probability analyses in visual and auditory decision tasks have provided especially strong evidence for this kind of temporally extended constraint. When stimuli such as noisy motion fields or contrast fluctuations are presented over time, behavioral choices can be regressed against the stimulus history to derive psychophysical kernels, revealing which moments of evidence most strongly influenced the final report. Parallel analyses applied to neural recordings show that spike trains in sensory areas reflect a similar weighting of stimulus time points, but they also exhibit late components that are modulated by the eventual choice beyond what is predicted from the stimulus alone. These late choice-related signals appear even in primary sensory cortex, indicating that decision-related feedback reshapes sensory representations retrospectively. From a time symmetry perspective, the same physiological circuits that initially encode sensory evidence later re-express activity patterns aligned with the chosen interpretation, as if the final decision were projecting constraints backward onto earlier processing stages.
Single-trial decoding studies using multivariate pattern analysis and population-level recordings further highlight temporally symmetric aspects of evidence integration. In tasks where ambiguous stimuli are presented for extended durations, classifiers trained on neural activity at different times can predict both the current sensory configuration and the subject’s ultimate report. Importantly, decoders trained on late-trial activity often perform better at predicting early perceptual states than decoders trained solely on early activity, indicating that late activity carries refined information about how earlier inputs have been interpreted. This pattern is particularly evident in higher-order association areas and prefrontal cortex, where population states near the time of response encode a distilled summary of the entire stimulus history. When this late information is used to decode what must have been inferred earlier in the trial, it effectively implements a form of retrodiction that mirrors the backward pass of bayesian inference, even though the underlying spikes occur in a strictly forward temporal order.
Neural recordings during tasks with delayed feedback or reward have supplied complementary evidence that learning-related signals propagate backward in time to influence representations associated with earlier events. Dopaminergic neurons in midbrain structures fire phasic bursts not only at the moment of unexpected reward but, with learning, at earlier cues that predict the reward. This temporal shifting of prediction error signals is mirrored in cortical and striatal circuits, where neurons begin to encode anticipated outcomes well before they occur. At the same time, synaptic plasticity mechanisms such as eligibility traces maintain a memory of recent pre- and postsynaptic activity, allowing delayed modulatory signals to retroactively assign credit to the appropriate synapses. Experimental manipulations that selectively disrupt dopaminergic signaling at specific time points show that reward feedback delivered at the end of a trial can modify responses to stimuli presented much earlier, consistent with learning rules that treat the full trial as a single object of optimization rather than as a series of independent steps.
Electrophysiological and imaging data from working memory and delay-period tasks provide another line of support. During tasks where a stimulus must be held in mind over a delay before a response, persistent or sequential activity patterns in prefrontal and parietal cortices maintain information about the sample. However, delay-period activity is also modulated by knowledge of upcoming demands, such as the type of response required or the timing of a go cue. When task contingencies are manipulated so that the same initial stimulus can lead to different future requirements, neural trajectories during the delay reorganize to align with the specific future action that will be needed. This context-dependent reshaping of delay activity implies that intermediate states are jointly constrained by the memory of past inputs and by prospective task demands. Analyses that reconstruct latent trajectories from population recordings often reveal that the geometry of neural activity during the delay is better explained as an interpolation between initial and final states than as a simple continuation of the immediate sensory response.
Studies of anticipatory activity in sensory cortices directly demonstrate how expectations about future events feed back to influence present representations. In vision, neurons in primary and higher areas can show pre-stimulus modulations of firing rates, orientation tuning, and receptive-field positions when a stimulus is expected at a particular time or location. Similar anticipatory modulations occur in auditory and somatosensory systems under temporal cueing paradigms. These pre-stimulus patterns often correlate with behavioral performance, such as detection sensitivity or reaction time, indicating that they embody useful predictions and priors about upcoming inputs. When the expected stimulus fails to appear, the same circuits exhibit mismatch responses or prediction error signals. The fact that neural activity before stimulus onset carries information about future inputs and that subsequent error responses reshape these expectations suggests a bidirectional interplay in which predictions established on the basis of learned regularities constrain how upcoming evidence will be processed, while mismatches retroactively update those predictive states.
Neuroimaging experiments in humans have provided coarse but compelling support for similar mechanisms at larger spatial scales. Functional MRI studies examining pattern completion, sequence prediction, and replay have shown that distributed cortical and hippocampal networks encode upcoming stimuli or actions several seconds in advance. For instance, in tasks where subjects learn structured sequences of visual or auditory elements, multivoxel patterns in hippocampus and association cortex begin to resemble the representation of the next item in the sequence before it is physically presented. When sequences are violated, error-related responses appear not only at the violation point but also at preceding time points when the predicted item would have occurred, reflecting a distributed encoding of expected trajectories through representational space. These observations align with models in which the brain maintains a probability distribution over entire future paths and updates it in light of mismatches, rather than merely forecasting the next step.
Direct evidence for temporally symmetric inference has come from experiments employing offline analysis techniques that explicitly use future data to reconstruct past neural states. For example, when latent dynamics are inferred from spiking activity using state-space models or dimensionality-reduction methods, algorithms that perform smoothing—integrating information from both past and future time bins—yield latent trajectories that more accurately predict behavior and sensory variables than those derived from purely forward filters. When these smoothed trajectories are examined, the inferred state at a given time often correlates more strongly with both earlier and later task variables than the raw observed activity does, suggesting that the underlying neural computation is organized around stable trajectories that reconcile constraints from the entire trial. While the smoothing itself is a post hoc analytic procedure, the success of such models in capturing behaviorally relevant structure indicates that the brain may be exploiting similar principles during real-time inference, with recurrent loops and feedback pathways providing biological substrates for approximating bidirectional message passing.
Experiments on perceptual multistability, such as binocular rivalry or ambiguous figures, also point toward time-symmetric components in the evolution of percepts. Neural recordings from visual cortex and higher-order areas reveal that activity not only tracks the currently dominant percept but also carries signatures of upcoming switches. In some cases, subtle changes in population activity precede a perceptual transition by hundreds of milliseconds, and similar patterns can be observed when the same percept is about to reemerge after a period of suppression. These anticipatory patterns imply that the system is exploring a constrained set of possible trajectories through representational space, with future perceptual states shaping the likelihood of particular metastable configurations at earlier times. When analyzed within the framework of energy landscapes, the transitions between percepts appear to follow paths that would be predicted by considering both the current state and the eventual attractor, echoing the kind of boundary-conditioned dynamics characteristic of time-symmetric formulations.
Sleep and memory consolidation studies provide additional support through the phenomenon of neural replay. During slow-wave sleep and quiet wakefulness, hippocampal place cells and associated cortical neurons reactivate sequences of activity that correspond to previously experienced trajectories through an environment. Critically, these replays occur not only in the original forward order but also in reverse. Reverse replay is particularly prominent following reward delivery, with sequences unfolding backward from the reward site to earlier locations along the path. This bidirectional replay is thought to support credit assignment, allowing information about delayed outcomes to be propagated back to preceding states. The existence of both forward and reverse sequences in the same circuitry indicates that the underlying connectivity and dynamics can support traversal of memory traces in either temporal direction, in line with models where the same internal structure supports both prediction and retrodiction over experienced trajectories.
Fine-grained analyses of dendritic and synaptic processes have begun to reveal microcircuit mechanisms that could underwrite such temporally extended computations. Experiments using two-photon calcium imaging and voltage-sensitive dyes show that distal dendritic compartments in pyramidal neurons can integrate feedback inputs carrying contextual and outcome-related information, while proximal compartments process more immediate feedforward drive. Nonlinear interactions between these compartments allow signals reflecting future-relevant constraints, such as expected reward or task rule, to modulate the impact of incoming sensory evidence in a temporally flexible manner. In some cases, the arrival of a late outcome signal can induce dendritic plateau potentials that retroactively strengthen or weaken synapses that were active earlier in the trial, effectively implementing a local form of time-symmetric credit assignment. These cellular-level observations are consistent with the idea that even within a single neuron, activity reflects a compromise between influences that are tied to different points along the behavioral timeline.
Neural oscillations and cross-frequency coupling also exhibit features suggestive of temporally symmetric processing across cycles. Studies using laminar recordings and magnetoencephalography have found that low-frequency rhythms, such as theta and alpha, can organize phases in which feedback or feedforward influence dominates, while higher-frequency gamma activity reflects local evidence accumulation. During tasks with predictable temporal structure, phase-amplitude coupling patterns shift as a function of both elapsed time and expected future events, indicating that oscillatory states anticipate upcoming task epochs. When these oscillatory dynamics are analyzed over multiple cycles, they form repeating motifs that align with both past and future task segments, as if the network were embedding the entire task structure into a recurrent temporal scaffold. This pattern is compatible with a view in which oscillations provide a carrier for alternating bouts of prediction and error correction, with each cycle approximating a step of bidirectional inference over a moving temporal window.
Experimental perturbations that selectively disrupt top-down or bottom-up pathways have helped dissociate causal from inferential time structure. Inactivation of higher-order areas or pharmacological disruption of feedback synapses often diminishes late-choice signals in early sensory cortex and reduces the alignment between sensory responses and eventual behavior, even when early feedforward responses remain largely intact. Conversely, perturbations that target ascending pathways can leave late top-down influences partly preserved, resulting in activity that reflects task expectations without accurate sensory grounding. These manipulations show that retrospective reshaping of sensory representations depends critically on intact feedback circuits, supporting the view that late constraints imposed by decisions, goals, or contextual inferences are fed back to earlier processing stages. The dependence of behaviorally relevant representations on both intact forward and backward pathways reinforces the idea that neural computation is implemented as a distributed negotiation between past-driven and future-oriented influences.
Converging evidence from cross-species studies suggests that these temporally extended and partially symmetric principles are not idiosyncratic to particular experimental paradigms but reflect a general organizational motif. From insect mushroom bodies involved in olfactory learning to mammalian hippocampal–prefrontal networks, neural systems consistently show patterns where activity at a given moment encodes information about both antecedent inputs and upcoming outcomes or actions. The ubiquity of phenomena such as anticipatory tuning, reverse replay, delayed credit assignment, and choice-dependent modulation of early sensory areas indicates that the brain routinely implements forms of evidence integration that span entire episodes. Interpreted through the lens of time symmetry, these findings suggest that neural processing is best understood not as a simple feedforward chain but as the evolution of states constrained by both historical inputs and future-oriented goals, with recurrent circuitry and plasticity mechanisms providing the substrate for approximating bidirectional inference within the unidirectional flow of physical time.
Implications for learning, prediction, and memory
Viewing learning through the lens of time symmetry highlights how the brain must reconcile information propagating from both past and future constraints when adjusting synaptic weights. Traditional formulations of plasticity emphasize causal relationships: presynaptic activity followed by postsynaptic depolarization leads to strengthening, whereas reversed timing leads to weakening. Yet these local timing rules operate within a broader context in which distal goals, delayed rewards, and long-range expectations determine whether a given activity pattern ultimately counts as successful. Synapses do not merely encode correlations with immediate consequences; they are embedded in credit-assignment structures that link early events to much later outcomes. In this sense, learning is less about storing snapshots of the past and more about shaping internal dynamics so that trajectories connecting past inputs to future goals become more probable. The same principles of bayesian inference that relate priors, likelihoods, and posteriors over full temporal sequences can be translated into synaptic updates that implicitly encode regularities stretching across both directions in time.
One concrete implication is that plasticity mechanisms must support a form of temporally extended evidence integration, where the impact of any single experience is filtered through estimates of how it fits into larger behavioral and environmental patterns. Eligibility traces and synaptic tags, for example, allow transient patterns of activity to remain “available” for modification until outcome signals arrive, effectively keeping a record of potential contributions to future success. When neuromodulatory bursts arrive at the end of a trial, they do not just reinforce states at that moment; they selectively modify synapses that were active earlier, reweighting the entire preceding trajectory in light of the eventual outcome. Learning rules that operate this way are functionally akin to gradient descent on an objective defined over complete paths. They use late information to revise the significance of past events, approximating a time-symmetric update even though the biophysical process of modulation unfolds strictly forward. The result is a learned internal model in which the same synaptic configuration supports both predicting future outcomes from early cues and retrospectively explaining which cues were most responsible for those outcomes.
In predictive processing frameworks, time symmetry becomes especially salient in how learning tunes the balance between prediction and priors on one hand and sensory-driven updates on the other. The brain is thought to minimize a long-term prediction error or free energy functional that spans entire episodes, not just instantaneous mismatches. Synaptic changes that reduce surprise at one time point are therefore evaluated with respect to their impact on the overall trajectory of errors across time. A synapse that improves accuracy early in a trial might nonetheless be weakened if it consistently leads the system into states that are hard to reconcile with later evidence or goals. Conversely, connections that initially appear suboptimal could be strengthened if they steer trajectories toward final states that are robust and behaviorally advantageous. This implies that learning is inherently future-sensitive: it cannot be fully described by statistics of past inputs alone but must also encode how present states participate in generating desirable future configurations of activity and behavior.
For prediction per se, time-symmetric perspectives stress that effective forecasting requires internal models that are also good at retrodiction—inferring plausible pasts from present states. When a network’s dynamics are shaped by an energy landscape or generative model defined over entire sequences, the same learned parameters that support next-step prediction also support reconstructing earlier hidden states, in line with the symmetry of bayesian inference. Biological systems can exploit this by using replay, rehearsal, and offline simulation to refine models in both temporal directions. During sleep or quiet wakefulness, forward and reverse replay of trajectories provide samples from the distribution of possible paths the organism experiences. Plasticity driven by these replays can improve the consistency of the internal model across time, ensuring that anticipatory activity before a stimulus and recollective activity after it fit into a coherent probabilistic structure. Prediction, in this view, is not merely about extrapolating a time series; it is about stabilizing a joint distribution over past, present, and future states, so that new evidence can be integrated with priors in a unified, time-symmetric manner.
Memory, too, acquires a different character when considered from a time-symmetric standpoint. Rather than treating memories as static records of past events, one can regard them as stored constraints on permissible future trajectories. A remembered episode does not merely describe what occurred; it encodes which sequences of states are likely to co-occur and how they culminate in outcomes such as rewards, dangers, or social consequences. When similar situations arise later, these stored trajectories act as priors over how the current episode might unfold, biasing perception and action toward paths that previously led to successful resolutions. At the same time, each new experience can retroactively modify existing memories, reinterpreting earlier events in light of later knowledge. This mutual influence is clearest in schema formation and memory reconsolidation, where newly learned information reshapes how prior experiences are encoded and retrieved. The resulting memory system functions less like a one-way archive and more like a living model that is continuously adjusted to maintain consistency across the temporal web of experiences.
Working memory and short-term maintenance of information can be reframed as the real-time expression of this trajectory-based modeling. Activity patterns during delay periods reflect not only what was just presented but also what operations must soon be performed—comparisons, transformations, or responses. When the same initial stimulus can lead to different future requirements, working memory representations diverge accordingly, even though the immediate sensory history is identical. This indicates that memory traces are not neutral placeholders; they are preconfigured to facilitate specific future computations. Under time-symmetric principles, the “content” of working memory is best understood as an intermediate state on a constrained path connecting past inputs with future actions. Learning then tunes recurrent connectivity so that these intermediate states occupy regions of the neural state space from which the appropriate future states are easily reachable while still being consistent with the initial evidence.
Long-term memory consolidation and systems-level reorganization of traces across hippocampus and neocortex can similarly be viewed as optimizing trajectories at multiple time scales. Shortly after an experience, hippocampal circuits may store detailed, episode-specific paths through state space. Over days to years, interactions between hippocampus and cortex gradually reshape cortical dynamics so that similar trajectories can be reproduced without explicit hippocampal guidance. Forward replay consolidates the predictive aspects of sequences, strengthening cortical pathways that lead from early cues to later outcomes. Reverse replay, in contrast, aids in propagating credit backward, reinforcing those cortical configurations that reliably precede successful end states. The interplay of these two directions of replay reduces redundancy, merges overlapping episodes, and yields semantic structures that support both efficient prediction and flexible recollection. Memory consolidation is thus not simply the transfer of information from one store to another but the progressive embedding of temporally extended constraints into the slow dynamics of cortical networks.
On the level of cognitive functions, time-symmetric neural computation suggests that learning, prediction, and memory are not separate modules but different modes of operation within a single inference machinery. Learning adjusts the parameters of the generative model so that its time-symmetric laws better match the statistics of experienced trajectories. Prediction corresponds to querying this model in a forward-biased way—conditioning on past and present to sample or infer future states that satisfy both external constraints and internal regularities. Memory retrieval corresponds to a complementary, backward-biased query—conditioning on current cues or outcomes to infer plausible past states that explain them. Both operations rely on the same underlying structure, and both are refined by the same plasticity processes that minimize long-run inconsistencies between modeled and experienced trajectories. The apparent asymmetries among learning, predicting, and remembering emerge primarily from which boundary conditions are emphasized at a given moment, not from fundamentally different computational principles.
These ideas also bear on how we interpret the subjective flow of time in cognition. From a formal perspective, the equations governing optimal inference and control over trajectories can often be written in forms that are nearly invariant under time reversal, aside from boundary terms capturing initial states, goals, and thermodynamic constraints. Yet our experience of remembering the past and anticipating the future is deeply asymmetric. This discrepancy can be understood by recognizing that while the underlying computational rules exhibit a degree of time symmetry, their real-time implementation is constrained by metabolic costs, noise, and limited memory. The brain cannot routinely perform full smoothing over arbitrarily long sequences, so it adopts strategies—such as forward-biased filtering during rapid online behavior and more balanced, bidirectional processing during offline periods—that approximate symmetric inference only under favorable conditions. Learning, prediction, and memory thus reflect a compromise between the mathematical symmetry of the ideal computations and the practical asymmetry imposed by a body that acts and senses in a world with a strong thermodynamic arrow of time.
Importantly, none of these considerations entail literal retrocausality; causes in the physical world still precede their effects. Time symmetry in this context refers instead to the structure of internal models and algorithms: the same learned relationships among events can be traversed in either temporal direction for the purpose of inference. When a future reward reshapes how past cues are stored, or when a later decision feeds back to modify representations of earlier stimuli, the underlying process is one of updating a joint model over entire episodes, not of signals traveling backward in physical time. Neural circuits exploit recurrent connectivity, feedback pathways, and plasticity mechanisms to approximate this joint modeling under real-world constraints. As a result, the systems that support learning, prediction, and memory come to embody principles that, at the algorithmic level, are much closer to time-symmetric optimization over trajectories than to a simple forward accumulation of associations.
