Time symmetry in recurrent computation rests on the idea that information processing should be describable in essentially the same way whether one follows activity forward or backward in time. In recurrent circuits, where feedback connections continually recycle activity, this symmetry becomes a powerful organizing principle. Instead of treating past inputs as causes and future states as mere consequences, a time-symmetric view allows both past and future constraints to shape the evolution of neural activity. Activity patterns can then be seen as trajectories through state space that are jointly determined by boundary conditions at multiple times, not only by initial conditions.
This perspective naturally aligns with the notion of the brain as a bayesian brain. In a Bayesian framework, internal states encode beliefs that combine priors with evidence, and these beliefs are continually updated to improve prediction and inference. A time-symmetric treatment of recurrent computation interprets these updates as the reconciliation of constraints coming from two temporal directions: priors and contextual structure that effectively propagate forward, and prediction errors or target information that can be thought of as propagating backward. The resulting dynamics approximate Bayesian inference by driving the system toward states that are simultaneously consistent with both past expectations and future outcomes.
In standard causal narratives of neural processing, signals flow from input to output in a strictly forward direction, and plasticity rules adjust synapses based on pre-before-post activity relationships. Time-symmetric principles, by contrast, suggest that changes in synaptic strengths should reflect correlations that are meaningful in both temporal directions. When a recurrent circuit settles into a low-error configuration, the pattern of pre- and postsynaptic activity carries information not only about what preceded it but also about what it predicts. Synaptic updates can then be interpreted as encoding regularities that are stable under time reversal: the same configuration that explains how the network got to its current state also explains how it will continue to behave.
This notion does not require literal retrocausality, in which future events physically affect the past. Instead, time symmetry is implemented at the level of descriptions and computations. A recurrent circuit can be designed so that the equations governing its evolution have a form that is invariant, or approximately invariant, under exchanging past and future indices. For example, energy-based models and certain attractor networks are defined by cost functions that do not privilege one temporal direction over the other. The network relaxes toward configurations that minimize a global quantity, and the same function can be used to explain both the forward relaxation and a hypothetical backward reconstruction of earlier states.
In such systems, error-driven signals can be interpreted symmetrically. From a forward-time perspective, deviations between predictions and observations generate errors that drive learning. From a backward-time perspective, the same error signals can be viewed as constraints that would have to be imposed on earlier states to ensure consistency with observed outcomes. A time-symmetric principle asserts that a single computational mechanism—implemented in the recurrent interactions—effectively performs both roles. This can be formalized by treating the network dynamics as gradient flows on a unified objective, where the same gradient governs both the forward inference trajectory and a notional backward correction trajectory.
Recurrent circuits that embody time-symmetric computation tend to blur the distinction between inference and learning. When the system receives an input and evolves over time, part of its trajectory corresponds to adjusting internal states to match the input, and part corresponds to adjusting effective synaptic influences to encode durable regularities. Under time-symmetric principles, these two adjustments can be viewed as different aspects of the same process: moving in state space toward configurations that are stable under both forward and backward reconstruction. In practical terms, this implies that the temporal unfolding of activity can carry embedded learning signals without the need for a separate forward pass for inference and backward pass for credit assignment.
From a probabilistic standpoint, time-symmetric recurrent computation can be interpreted as implementing joint constraints over entire trajectories rather than stepwise updates conditioned only on the immediate past. Instead of specifying transition probabilities that depend solely on previous states, one can define a path probability or trajectory energy that assigns a score to the whole sequence of neural states. The dynamics of the recurrent network then perform an implicit optimization of this path score. Crucially, such a formulation treats intermediate states as latent variables to be inferred given observations at multiple time points, which resonates with ideas from smoothing in state-space models and with the bayesian brain hypothesis that perception involves inferring hidden causes over time.
Another way to view time-symmetric principles is through the lens of prediction and retrodiction. Prediction uses current states to estimate future ones; retrodiction uses current information to refine beliefs about the past. In classical signal processing, these two problems are solved by different algorithms. In a time-symmetric recurrent circuit, however, the same dynamical rules can support both, depending on how boundary conditions are imposed. If inputs specify earlier states and targets specify later states, the network’s relaxation integrates both sets of constraints. The resulting internal trajectory encodes a compromise between forward prediction and backward retrodiction, which in turn shapes synaptic adaptation.
Time symmetry also imposes structure on the way recurrent connectivity is organized. For the dynamics to admit a time-reversal description, certain balances must hold between excitation and inhibition, feedforward and feedback pathways, and the distribution of delays across loops. These balances ensure that information about constraint violations—differences between what the current state predicts and what boundary conditions require—can propagate through the network without distortion in either temporal direction. When such balances are in place, local interactions can implement global credit assignment because the same pathways that carry predictive signals forward also carry corrective influences backward in an effectively symmetric manner.
Plasticity mechanisms compatible with time-symmetric principles often depend on correlations that are integrated over full trajectories rather than on instantaneous pre-post spike timing alone. For instance, eligibility traces can preserve information about past activity until a complementary signal, such as a delayed error or modulatory pulse, arrives. In a time-symmetric framework, these traces can be viewed as storing the forward-time component of the correlation, while the later modulatory signal embodies backward-time constraints derived from outcomes. The interaction between the two realizes a synaptic update that is invariant, in an abstract sense, under interchanging the order of cause and effect along the trajectory.
By emphasizing invariance under time reversal, time-symmetric principles promote a unification of neural computation across tasks that span perception, memory, and control. Recurrent circuits that obey such principles can reuse the same connectivity patterns to solve tasks that are ordinarily framed as distinct: filtering versus smoothing, recognition versus generation, or decoding versus encoding. The common thread is that all of these tasks can be cast as finding trajectories that satisfy constraints at multiple times, with the recurrent dynamics providing a universal engine for enforcing consistency. This unification hints at why biological recurrent networks may be so heavily recurrent and why their apparent complexity can, at a deeper level, be governed by relatively simple symmetric rules.
Formal framework for bidirectional learning dynamics
A formal description of bidirectional learning dynamics begins by treating the recurrent circuit as a dynamical system defined on a joint space of neural states and synaptic parameters. Let the instantaneous neural state be denoted by a vector (x_t), the synaptic parameters by (W), and the inputs and targets by (u_t) and (y_t), respectively. Instead of specifying separate mechanisms for forward inference and backward error propagation, the dynamics are derived from a single scalar functional that scores entire trajectories ({x_t}_{t=0}^T) and parameters (W). This trajectory functional plays the role of an energy or action: it is defined over the full temporal path, incorporates both priors and data likelihood, and contains explicit terms that encode the mismatch between the circuit’s own predictions and externally imposed boundary conditions at multiple time points.
Within this framework, the evolution of neural activity is described as a gradient flow on the trajectory functional with respect to the states (x_t). At any moment, the local update of (x_t) depends on the derivative of the global functional evaluated at that time, which implicitly incorporates both past and future constraints. When boundary conditions are imposed at the beginning and end of a trial—such as a clamped input at (t = 0) and a desired output at (t = T)—the gradient at intermediate times contains information flowing backward from the final constraint and forward from the initial one. The same recurrent connections that mediate predictive interactions among units also carry this constraint information, so that the effective forces on the state variables are symmetric under the notional reversal of time.
To make this symmetry explicit, it is useful to write the trajectory functional as a sum of three components: a prior term on trajectories that encodes structural preferences or regularities of the circuit, a data-consistency term that measures how well the generated activity fits observed inputs and targets, and a complexity term on synaptic parameters that biases the network toward simpler solutions. The prior term defines which trajectories are considered typical even in the absence of external signals, often enforcing smoothness, stability, or attractor structure. The data-consistency term couples the internal trajectory to boundary conditions at multiple times, not only at the outset. Crucially, both components are defined in such a way that their contribution to the gradient at a time point (t) can be decomposed into influences that propagate from earlier and later times, enabling a bidirectional interpretation of the resulting dynamics.
State updates in this setting can be viewed as implementing a form of temporal smoothing rather than simple filtering. In filtering, estimates of the current state depend only on past observations; in smoothing, they depend on the full set of observations, past and future. Bidirectional learning dynamics embed smoothing directly into the recurrent evolution of the network: as activity relaxes, it assimilates information about future constraints via feedback connections and uses this information to refine earlier parts of the internal trajectory. This is mathematically captured by the fact that the stationary points of the trajectory functional satisfy a set of coupled equations linking (x_t) to both (x_{t-1}) and (x_{t+1}), so that each state is adjusted to simultaneously reduce inconsistencies with its causal history and its expected consequences.
Synaptic plasticity enters the formalism as gradient descent on the same trajectory functional with respect to (W). Instead of computing gradients by running an explicit backward pass, the idea is that the time-extended evolution of the recurrent circuit itself encodes the necessary sensitivity information. During a learning episode, the network is exposed to inputs and, possibly at later times, to target signals. Its states evolve under the influence of both feedforward drive and feedback constraints until a stationary or quasi-stationary trajectory is reached. The correlations accumulated along this trajectory, often stored in eligibility traces at the synapses, are then combined with modulatory signals that summarize the mismatch between predicted and imposed boundary conditions. The resulting synaptic update can be shown to approximate the gradient of the overall functional, and because this functional is time-symmetric by construction, the update rule inherits a bidirectional interpretation.
To formalize how eligibility traces support this symmetry, one can introduce an auxiliary variable at each synapse that integrates products of pre- and postsynaptic activity over time. These variables obey local differential equations driven only by the activity of the neurons they connect, independent of delayed error information. Later, when an error-related modulatory signal becomes available—perhaps when a reward or target is revealed—the synaptic weight is adjusted by multiplying the stored eligibility with the modulatory signal. From a forward-time viewpoint, the eligibility trace captures how the synapse influenced the trajectory that led to the current error. From a backward-time viewpoint, the modulatory signal can be interpreted as specifying how future boundary conditions project backward through the circuit. Combining the two yields a plasticity rule that matches the gradient of a time-symmetric objective without requiring explicit nonlocal credit assignment.
Mathematically, the equivalence between this local plasticity rule and gradient descent on the trajectory functional relies on a correspondence between the network’s relaxation dynamics and an implicit adjoint system. In conventional gradient-based learning, the adjoint variables, or error signals, are propagated backward in time through the network using the transpose of the Jacobian of the forward dynamics. In the bidirectional framework, recurrent circuits are arranged so that the physical activity of certain units, or particular components of their activity, plays the role of these adjoint variables. Their evolution is governed by the same or closely related connectivity as the forward activity, ensuring that the path taken by these error-like signals is the time reverse, in an abstract sense, of the path taken by predictive signals. Thus, the same recurrent matrix that supports prediction also supports credit assignment, and the apparent separation between inference and learning reduces to different projections of a single, unified dynamical process.
This unification can be expressed compactly by augmenting the state space to include both “primal” variables, representing the neural activities that encode predictions or beliefs, and “dual” variables, representing constraint forces that enforce agreement with boundary conditions. The joint evolution of primal and dual variables is governed by a Hamiltonian or Lagrangian structure derived from the trajectory functional. Time reversal in this setting maps primal variables at time (t) to corresponding configurations at time (T – t) and dual variables to their negatives, leaving the underlying equations invariant. Learning then corresponds to adjusting parameters so that the observed trajectories of these augmented states minimize the action. Because the equations are time-symmetric, a given synaptic change can be interpreted equally well as improving forward prediction accuracy or backward reconstruction fidelity.
From a probabilistic perspective, this formal framework mirrors the structure of variational inference in state-space models, but instantiated in physical recurrent circuits. The trajectory functional is analogous to a variational free energy that upper-bounds the negative log evidence of observed sequences. Minimizing it with respect to neural trajectories performs approximate inference over latent states; minimizing it with respect to synapses performs learning of generative and recognition parameters. In a time-symmetric instantiation, the recognition dynamics—how the circuit updates its internal states in response to new data—and the learning dynamics—how synapses change to better encode regularities—are generated by the same underlying variational principle. This correspondence tightens the link between the bayesian brain hypothesis and concrete models of recurrent circuits whose activity trajectories and plasticity rules are constrained by time-reversal symmetry.
Importantly, the formalism also clarifies what time symmetry does not imply. It does not require any form of physical retrocausality; instead, it demands that the mathematical description of both inference and learning be expressible in a way that is invariant, or nearly invariant, under swapping the roles of past and future in the trajectory functional. Deviations from exact symmetry are expected when the environment is strongly irreversible or when the circuit incorporates directional constraints such as conduction delays and asymmetric noise. Nonetheless, by organizing the learning dynamics around a nearly symmetric objective defined over full trajectories, one can design recurrent circuits whose local operations approximate the ideal bidirectional gradients. This approximation suffices to endow the system with the ability to use future outcomes to refine representations of earlier states and to share computational resources between prediction, retrodiction, and long-term synaptic modification.
Implementation of time-reversed plasticity rules
Implementing time-reversed plasticity rules in recurrent circuits requires translating the abstract, trajectory-based gradients of a time-symmetric objective into locally computable updates at each synapse. The key idea is that synapses should respond not only to instantaneous pre- and postsynaptic activity, but to a structured decomposition of that activity into forward- and backward-propagating components. Forward components reflect the influence of inputs and priors, whereas backward components encode constraints derived from outcomes or targets. A time-reversed rule then emerges when weight changes depend on correlations between these two components, such that exchanging their temporal roles leaves the update invariant in form. This turns the synapse into a device that effectively “listens” simultaneously to how its activity participates in predicting the future and how it would need to change to better retrodict the past.
One practical route to this implementation is to explicitly split each neuron’s activity into two channels: a predictive channel that carries the usual recurrent computation and a corrective channel that carries error-like signals. The predictive channel evolves forward in time under the influence of inputs, recurrent feedback, and internal noise; the corrective channel is driven by discrepancies between network outputs and desired boundary conditions and tends to propagate “backward” through the same or mirrored connectivity. Synaptic plasticity then depends on a product of the pre-neuron’s predictive activity and the post-neuron’s corrective activity, integrated over time and modulated by a global signal that captures the overall quality of the trajectory. Because both channels use similar recurrent pathways, inverting the direction of time approximately swaps their roles: the former becomes an error carrier and the latter becomes predictive. Designing the plasticity rule to be symmetric under this swap enforces time-reversal invariance at the synaptic level.
Eligibility traces provide the substrate for this symmetry. Each synapse maintains a hidden variable that accumulates a filtered history of pre–post co-activity during the predictive phase. This accumulation can be implemented as a simple leaky integrator driven by products of pre- and postsynaptic firing rates or spike trains. Crucially, the integration window must be long enough to span the temporal separation between the onset of predictive activity and the eventual arrival of outcome-dependent information. When a delayed teaching or reward signal arrives—effectively encoding backward-time constraints—it multiplies the stored eligibility to produce an actual weight change. Under a time-symmetric interpretation, the build-up of eligibility corresponds to the forward-time contribution to the gradient, while the delayed modulatory pulse embodies the backward-time contribution; their interaction reconstructs the full trajectory gradient without an explicit backward pass.
To more closely approximate exact time-reversal symmetry, the temporal filtering in eligibility traces can be shaped to mirror the statistics of constraint propagation in the network. For instance, if recurrent loops introduce specific delays and reverberation times, the decay constants of the traces can be tuned so that the eligibility available at the moment of outcome revelation is proportional to the sensitivity of the outcome to that synapse’s earlier activity. In continuous-time formulations, this tuning corresponds to matching the kernel of the eligibility filter to the Green’s function of the forward dynamics. Under time reversal, the same kernel describes how a perturbation at the time of the outcome would propagate backward toward earlier states. A plasticity rule that weights eligibility by outcome-related signals using this matched kernel thereby becomes invariant under reversing the arrow of time, at least within the approximations of the model.
An alternative but related implementation uses phase-based or oscillatory codes to differentiate forward and backward influences. In such schemes, pre- and postsynaptic activities are decomposed into components locked to different phases of an ongoing rhythm, with one phase band preferentially carrying predictive signals and another carrying retrodictive or error signals. Synaptic plasticity is then gated not only by the amplitude of activity but also by its phase relationship: correlations between a presynaptic predictive phase and a postsynaptic error phase drive weight changes, while correlations in the opposite combination may be suppressed or assigned opposite sign. If the phase relationships are arranged so that advancing or rewinding time corresponds to swapping these roles, then the same underlying spike sequences, viewed under time reversal, induce equivalent but time-reversed patterns of plasticity. This implements a form of time-symmetric learning in which oscillatory structure encodes the implicit backward pass.
Spike-timing-dependent plasticity (STDP) can also be reformulated to support time-symmetric rules by broadening and balancing its temporal window. Standard STDP often privileges causal pre-before-post pairings, leading to asymmetric learning that encodes temporal precedence. A time-reversed variant, by contrast, uses a biphasic window that assigns comparable but sign-structured weight to pre-before-post and post-before-pre pairings across a longer timescale. The magnitude and sign of synaptic change then depend on whether the pattern of spikes is consistent with both forward and backward interpretations of the underlying temporal structure. When integrated across entire trajectories and combined with outcome-dependent modulation, such a balanced STDP rule approximates the gradient of a symmetric trajectory cost: it strengthens synapses that participate in patterns reproducible under time reversal and weakens those that contribute to temporally inconsistent predictions.
In networks that explicitly represent both primal and dual variables, time-reversed plasticity can be implemented through cross-coupling between these two populations. Primal units encode beliefs or predictions about latent causes, while dual units encode constraint forces or error signals. The recurrent circuits are arranged so that primal-to-primal and dual-to-dual connections share parameters or are constrained to be symmetric, and primal-to-dual connections mirror dual-to-primal ones. Synaptic updates are then computed by correlating primal activity on one side of a synapse with dual activity on the other side. Because the dual variables follow dynamics equivalent to an adjoint system that runs backward in time with respect to the primal trajectory, the resulting weight change corresponds to a product of forward and backward sensitivities. Under an abstract time reversal that exchanges primal and dual roles and flips the direction of time, the plasticity rule retains its form, making it a concrete realization of time-reversed learning.
Energy-based implementations offer another constructive example. Consider a recurrent network whose dynamics perform gradient descent on a scalar energy function that depends on both neural states and synaptic weights. During a “free” phase, the network receives inputs but no target constraints and relaxes toward a low-energy configuration; during a “nudged” phase, targets or outcome constraints are imposed, slightly modifying the energy landscape and causing the state to relax to a different configuration. A time-symmetric plasticity rule compares synaptic correlations between these two phases, updating weights in proportion to the difference. From a forward-time view, the nudged phase injects future information that drives correction of the earlier free-phase prediction; from a backward-time view, the free phase can be interpreted as a reconstruction attempt given final constraints. The differential correlation rule is invariant under exchanging which phase is considered “before” or “after,” provided the nudging is small and symmetric, yielding an effectively time-reversal-consistent learning procedure.
Implementations that aim to be biologically plausible often rely on global modulatory signals, such as neuromodulator concentrations or dendritic plateau potentials, to encode outcome- or reward-related information. In a time-symmetric scheme, these modulators are not simple scalar rewards, but proxies for the mismatch between actual and expected boundary conditions of entire trajectories. A neuromodulatory pulse arriving at the end of a sequence can encode whether the realized trajectory satisfied both past-driven priors and future-driven goals. Synaptic plasticity then arises from the interaction between this global signal and locally stored eligibility traces. Because the modulatory signal is broadcast to wide areas and because the eligibility traces preserve a record of earlier activity, the resulting updates approximate the global trajectory gradient in a purely local manner. Conceptually reversing time reinterprets the modulatory pulse as a prior constraint and the eligibility traces as records of how later states depend on earlier ones, leaving the multiplicative update unchanged.
To make these mechanisms robust in noisy, high-dimensional settings, additional regularization can be embedded directly into the plasticity rule. For example, synapses can be endowed with their own slow dynamics that pull weights toward baseline values or low-complexity structures unless supported by consistent, time-symmetric correlations. This prevents runaway growth from transient coincidences and biases learning toward patterns that recur across many trajectories and directions of time. Likewise, local normalization mechanisms—such as synaptic scaling or competition among incoming synapses—can be applied to the magnitude of eligibility traces and weight changes, ensuring that the contributions of forward and backward influences remain balanced. Under time reversal, these normalization processes act equivalently, preserving the symmetry of the learning dynamics.
In hardware or algorithmic implementations, enforcing time-reversed plasticity rules often requires careful coordination between how recurrent activity is updated and how plasticity is triggered. For discretized simulations, one can alternate update steps that compute forward predictions with steps that compute backward corrections, while maintaining eligibility buffers that bridge the two. For analog neuromorphic systems, circuit components such as capacitors and transconductance amplifiers can be configured so that the same physical pathways support both the accumulation of predictive influences and the propagation of corrective currents when boundary conditions change. In both cases, the design principle is that the path used by error-like signals to reach a synapse should be, in an appropriate sense, the time reverse of the path used by predictive signals. When this condition holds and plasticity depends on their joint correlations, the network naturally implements learning rules that treat prediction and retrodiction on equal footing, embodying the core idea of time-symmetric adaptation in recurrent circuits.
Applications to sequence learning and prediction
Sequence learning provides a natural arena in which time-symmetric principles reveal their full power, because temporal structure is both the object of learning and the substrate on which computation unfolds. In conventional approaches, a network observes an input stream, transforms it through recurrent connections, and generates a prediction of the next element or a label for the entire sequence. Learning proceeds by computing an error at the end of the sequence and propagating it backward through time using backpropagation through time. In a time-symmetric framework, by contrast, the same recurrent circuits that generate predictions also act as physical substrates for propagating information about future mismatches back toward earlier states. The trajectory of activity across time is shaped simultaneously by forward influences, such as sensory drive and priors over sequences, and backward influences, such as target constraints or surprise about upcoming items. Synaptic plasticity then encodes correlations that are coherent with both directions, enabling the network to internalize the global structure of sequences rather than only local next-step dependencies.
Consider a simple sequence prediction task in which the network must anticipate the next symbol in a stream based on context several steps in the past. A purely forward model treats each step as a conditional prediction problem, estimating (p(s_{t+1} mid s_{le t})). A time-symmetric model, however, implicitly approximates the joint distribution over entire sequences (p(s_{0:T})) and performs inference over trajectories of hidden states that must be compatible with the full observed context. When an unexpected symbol occurs near the end of a sequence, the mismatch does not merely adjust weights based on local activity at that time. Instead, constraint signals propagate backward through the recurrent connectivity, reshaping earlier parts of the internal trajectory so that they become more consistent with the final outcome. During repeated exposure to similar streams, synapses update according to these bidirectional influences, gradually learning internal dynamics that favor trajectories consistent with both early context and typical later continuations. The result is a model that captures long-range dependencies without the brittleness that often arises when credit assignment is forced to rely on a separate, explicitly backward computational pass.
From a probabilistic standpoint, this approach corresponds to moving from online filtering to temporal smoothing in the way sequences are represented. Filtering attempts to maintain, at each time point, a belief about the current latent state conditioned on all past observations. Smoothing refines that belief using data from both past and future. Time-symmetric recurrent computation effectively implements a smoothed representation by allowing later constraints to refine the interpretation of earlier events. For example, in language modeling, an ambiguous word early in a sentence may have multiple plausible interpretations. As later words arrive and disambiguate the intended meaning, constraint-like activity flows backward through the circuit, reconfiguring the latent representation of the earlier segment. Plasticity rules that depend on the overlap between these revised internal states and the observed sequences then reinforce synaptic configurations that support globally consistent interpretations. In this way, the network learns sequence encodings that are optimized not just for immediate prediction, but for joint coherence with eventual outcomes.
Applications to supervised sequence labeling tasks provide another concrete illustration. In tasks such as phoneme recognition from acoustic waveforms or part-of-speech tagging from text, labels may be available only at selected time points or only at the end of the sequence. A time-symmetric network embeds these labels as boundary conditions that anchor the trajectory of internal states at specific times. The recurrent dynamics then interpolate between labeled and unlabeled segments, propagating constraint information both forward and backward. Synaptic learning depends on correlations integrated across the entire trajectory, so that unlabeled intervals that systematically lead into or out of particular labels acquire characteristic representations. Over training, the circuit learns to allocate internal states along the sequence that predict not just the immediate next step but also the later labels that are most compatible with the entire observed context. This yields more robust segmentation and classification, especially in cases where local cues are weak and only the global pattern identifies the correct label.
Time-symmetric principles are particularly powerful in multiscale sequence learning, where structure occurs at multiple nested temporal resolutions. Real-world sequences, such as music, speech, or motor behaviors, contain motifs and subroutines that are reused across contexts, as well as slower-varying plans or intentions that shape how motifs are combined. In a time-symmetric recurrent architecture, slower timescale units can encode priors over entire subsequences or episodes, while faster units encode detailed transitions. When an unexpected event occurs at the fast timescale, backward-propagating constraint signals can adjust not only the immediate transitions but also the slower contextual states, effectively updating the inferred “plan” that best explains the full trajectory. Subsequent learning then strengthens synapses associated with context–motif pairings that remain stable under such bidirectional adjustments. The network thereby acquires hierarchical sequence knowledge in which both fine-grained transitions and coarse-grained structure are tuned to minimize inconsistencies over full temporal segments.
Reinforcement learning in sequential decision-making tasks also benefits from time-symmetric treatment. In standard formulations, agents use prediction errors, such as temporal-difference signals, to update value estimates and policies, often propagating reward information backward through an eligibility trace mechanism. Time-symmetric recurrent circuits generalize this idea by treating entire state–action trajectories as objects of inference under a trajectory-level objective. Rewards or costs imposed at the end of an episode act as final boundary conditions that shape the entire relaxation process. Dual-like units or error channels carry signals representing the discrepancy between realized returns and those expected under current internal models, and these signals flow backward along the same or closely mirrored pathways used to generate action predictions. Synaptic updates then depend on the overlap between forward-rolling belief or policy signals and backward-rolling return-related signals. This yields value and policy updates that are naturally aligned with the full temporal extent of behavior, potentially stabilizing credit assignment in long-horizon tasks where sparse rewards challenge purely forward approaches.
In model-based reinforcement learning, where agents attempt to learn an internal dynamical model of the environment, time-symmetric learning casts the problem as discovering latent dynamics that can generate and retrodict observed trajectories with equal facility. The recurrent network learns a generative process that, when run forward, produces plausible future state sequences given current context, and when constrained by later outcomes, can reconstruct the most likely past trajectory that led to them. During training, observed transitions and rewards specify partial segments of trajectories; the network’s internal states evolve to fit these segments while remaining consistent with its priors. Deviations between predicted and actual sequences generate error-like influences that are propagated both forward and backward in time through the model’s recurrent connectivity. Plasticity driven by these bidirectional signals tunes the synapses so that the same internal dynamics become adequate for planning future trajectories and for explaining past ones. This tight coupling between planning and explanation can improve sample efficiency, because each experience contributes information to both forward prediction and backward credit assignment.
Sequence completion and inpainting tasks provide a direct testbed for time-symmetric recurrent computation. Here, the system is presented with fragments of a temporal pattern, such as missing segments in a sound or occluded frames in a video, and must reconstruct the absent portions. In a purely feedforward setting, this often requires specialized architectures that map observed context to a predicted continuation. In a time-symmetric recurrent setting, the observed fragments are simply imposed as constraints at their respective time points, and the network is allowed to relax. Information flows from the known segments in both directions, sculpting the activity at missing intervals so that the overall trajectory satisfies the global dynamics and all boundary conditions. During learning, synaptic changes reinforce those internal dynamics that reliably allow the circuit to fill such gaps while remaining consistent with overall patterns. Because the same dynamics are used for completion in either temporal direction, the model can perform both forward extrapolation and backward interpolation without architectural changes.
Temporal pattern recognition under uncertainty is another area where time-symmetric approaches excel. Natural sensory streams are often noisy and incomplete, with ambiguous or partially missing cues. A bayesian brain–like interpretation views recognition as inferring latent causes that best explain the totality of observations. Time-symmetric recurrent circuits make this view operational: latent representations at each time step are adjusted by a combination of bottom-up evidence, top-down priors, and lateral consistency constraints across time. When later observations contradict early interpretations, backward-propagated constraints revise earlier latent states; downstream recognition decisions are based on these revised trajectories rather than on the initial, purely forward estimates. Plasticity that integrates over the resulting trajectories strengthens synapses that support latent representations stable to such bidirectional corrections. This naturally leads to sequence encoders that are robust to local noise and can exploit future evidence to interpret earlier ambiguous segments.
In predictive coding formulations, each layer of a hierarchical recurrent network attempts to predict the activity of the layer below over time, while dedicated error units encode deviations between predictions and actual input. Time-symmetric learning introduces a complementary role for “retrodictive coding,” in which activity at a given time is influenced by how well it can be reconciled with later prediction errors. For instance, in a visual sequence, a sudden change in motion direction late in a clip can propagate backward as a wave of error signals, leading the network to reinterpret earlier frames as part of a more complex motion pattern (such as a turn rather than a straight trajectory). Synaptic updates that depend jointly on the predictive activity and these delayed error signals functionally approximate the gradient of a trajectory-level prediction objective. Because the same generative pathways are used for both prediction and the propagation of backward constraints, the learning process treats future events as sources of information about how earlier predictions should have been formed, without invoking literal retrocausality at the physical level.
In sensorimotor sequence learning, as in reaching, locomotion, or speech articulation, actions and sensory consequences form tightly coupled temporal patterns. Time-symmetric recurrent circuits can be used to learn internal forward models that predict sensory outcomes from motor commands, as well as inverse models that infer commands from desired outcomes, using a single recurrent substrate. During practice, the agent executes motor sequences and observes the resulting sensory feedback; discrepancies between expected and actual trajectories of proprioceptive and exteroceptive signals generate constraint-like activity that spreads across the network. These signals not only adjust synapses responsible for immediate motor-to-sensory mappings but also refine representations of earlier command segments that set up the context for later outcomes. Over time, plasticity guided by these bidirectional signals yields internal dynamics that support smooth movement generation forward in time, and also the reconstruction or refinement of motor commands when desired end states are specified as boundary conditions. This dual capability underlies flexible motor planning, correction of ongoing actions, and learning from delayed feedback.
Applications to continual and lifelong learning highlight an important advantage of time-symmetric approaches for sequence processing. In non-stationary environments, patterns that were once predictive may later become misleading, and new structures may emerge that reinterpret historical data. Time-symmetric recurrent circuits allow later experiences to retroactively influence the interpretation of earlier ones encoded in synaptic structure. When confronted with a new regularity that sheds light on previously unexplained variations, backward-propagating constraint signals can selectively adjust weights that participated in earlier, now-recognized patterns, without globally erasing prior knowledge. Because plasticity is driven by correlations across entire trajectories, including both historical priors embedded in the dynamics and new future-facing constraints, the network can gradually reshape its internal sequence model to accommodate evolving structure while preserving components that remain consistent under both past and future evidence. This capacity to let future experiences refine the meaning of past activity, implemented without explicit storage of raw histories, is a central practical benefit of embedding time-symmetric learning directly into recurrent circuits for real-world sequence prediction and interpretation tasks.
Implications for biologically plausible neural circuits
Viewing real neural tissue through the lens of time-symmetric learning forces a re-evaluation of what constitutes a biologically plausible mechanism for credit assignment. Many standard machine learning algorithms rely on explicit backward passes, nonlocal gradient information, or precise weight transport, features that are difficult to reconcile with the anatomy and physiology of cortical networks. Time-symmetric principles instead suggest that the very same recurrent circuits that carry predictive activity may also carry the effective “backward” influence of outcomes, provided that activity and plasticity are organized around trajectory-level constraints rather than stepwise causal chains. In this view, backpropagation is not a separate operation layered on top of neural computation, but an emergent property of how distributed populations relax toward states consistent with both past inputs and future targets.
This perspective dovetails naturally with the bayesian brain hypothesis. If cortical circuits are performing approximate inference over latent causes of sensory streams, then they must combine priors with evidence distributed across time. A time-symmetric implementation treats priors as constraints that propagate “forward” through generative pathways and prediction errors as constraints that, effectively, propagate “backward.” Biologically, these two directions need not be encoded in distinct anatomical pathways. Instead, they may be multiplexed in the same connectivity via different cell types, laminar circuits, or temporal phases of activity. The net effect is that neural states at any time embody a compromise between forward-propagating expectations and backward-propagating corrections, with synaptic plasticity capturing correlations that remain stable under this bidirectional flow.
Cortical microcircuit motifs offer candidate substrates for such bidirectional dynamics. The canonical layered structure, with reciprocal connections between superficial and deep layers, already supports cycles of feedforward and feedback signaling. Superficial pyramidal neurons receive thalamic and lower-level cortical input, while deep pyramidal neurons project feedback to earlier areas and subcortical structures. Within a time-symmetric framework, one can interpret superficial activity as emphasizing forward-looking prediction and deep activity as emphasizing constraint or error-like information, even though both populations are embedded in the same recurrent loops. Interlaminar and intralaminar connections then orchestrate relaxation processes in which predictions and corrections continually interact until a locally consistent pattern emerges, effectively implementing a form of temporal smoothing at the circuit level.
Inhibitory interneurons further enrich this picture. Distinct interneuron classes, such as parvalbumin-positive, somatostatin-positive, and vasoactive intestinal peptide–positive cells, differentially target somatic, dendritic, and other interneuron compartments. These diverse circuits can regulate the gain and timing of excitation, shaping when and how information about priors and outcomes is integrated. For instance, dendrite-targeting interneurons may gate the impact of top-down feedback onto pyramidal cells, effectively controlling the strength of backward-propagating constraints. Somatic-targeting interneurons can synchronize local assemblies, aligning phases of activity during which predictive and error-like components are most strongly correlated. The resulting balance of excitation and inhibition, tuned by experience-dependent plasticity, can stabilize recurrent dynamics while still allowing the network to reconfigure trajectories in response to delayed feedback.
Dendritic computation is especially important for realizing local approximations to time-symmetric learning. Pyramidal neurons exhibit complex, nonlinear integration across basal and apical dendrites, with apical tufts receiving substantial feedback and modulatory input. One biologically plausible interpretation is that basal dendrites predominantly integrate forward-flowing sensory and lateral information, whereas apical dendrites carry backward-flowing prediction errors or goals. When coincident patterns of activity arrive in these compartments within an appropriate temporal window, active dendritic events such as NMDA spikes or Ca2+ plateaus are triggered, which in turn drive plasticity at nearby synapses. The coincidence of basal “prediction” and apical “constraint” inputs effectively computes a local proxy of the product between forward and backward signals, making synaptic updates consistent with gradient-like time-symmetric plasticity without explicit nonlocal computations.
Neuromodulatory systems provide another ingredient required for trajectory-level learning in biological circuits. Dopamine, serotonin, acetylcholine, and norepinephrine are released in response to salient events, rewards, uncertainty, and violations of expectation, respectively. These neuromodulators can be interpreted as broadcasting global signals that summarize the quality of recent trajectories relative to organismal goals. In a time-symmetric account, neuromodulatory pulses arriving at key points in a behavior act as boundary-condition indicators: they signal whether the trajectory that just unfolded was better or worse than expected. Local synapses, having maintained eligibility traces of recent co-activity, convert these global pulses into weight changes. Because the eligibility traces reflect forward-time contributions and the modulatory signals reflect backward-time evaluations, their interaction generates updates that depend on both directions in time, approximating a symmetric trajectory-gradient using only local information.
Empirical evidence for such eligibility-based learning has been accumulating across brain areas. In corticostriatal circuits, spike-timing–dependent eligibility traces can persist for hundreds of milliseconds or longer before dopaminergic signals determine the sign and magnitude of plasticity. In cerebellar circuits, complex spikes in Purkinje cells, driven by climbing fiber input that encodes error-like signals, interact with granule cell activity history to induce long-term depression or potentiation. These mechanisms align well with time-symmetric principles: they separate the storage of forward-propagating influence (what the neuron did) from the arrival of backward-propagating evaluation (how that action affected outcomes). When combined across many synapses within recurrent circuits, such rules allow the network to internalize regularities that shape entire trajectories of activity and behavior, not just instantaneous responses.
Oscillations and phase coding in the brain also support a biologically grounded implementation of bidirectional dynamics. Rhythms in the theta, alpha, beta, and gamma ranges coordinate activity across local and distributed networks. Time-symmetric theories suggest that different phases of an oscillation could preferentially carry predictive versus error-like signals, even within the same anatomical pathways. For example, in hippocampal–entorhinal circuits, phase precession and phase-of-firing codes allow neurons to represent sequences of locations within a single theta cycle. One possible interpretation is that early phases represent forward trajectory prediction, while later phases incorporate feedback from actual sensory or proprioceptive outcomes. Synaptic plasticity that is sensitive to phase relationships, as observed in several experiments, can then selectively reinforce correlations consistent with both forward and backward sequence structure, providing a naturally oscillatory substrate for time-symmetric learning.
Hippocampal replay phenomena offer a particularly vivid example of how the brain may approximate time-reversed processing. During rest and sleep, place cells spontaneously reactivate sequences that mirror those experienced during wakefulness. These replays can occur in both forward and reverse order relative to behavior. Forward replay is often associated with planning or evaluating future paths, whereas reverse replay has been linked to consolidating reward information into earlier parts of a trajectory. From a time-symmetric viewpoint, replay provides off-line opportunities for recurrent circuits to explore both causal and retrocausal interpretations of experienced sequences, without invoking literal retrocausality at the physical level. Synaptic changes driven by patterns that are stable under both forward and reverse replay naturally implement learning rules that respect trajectory-level symmetry, strengthening internal models that support accurate prediction and efficient retrodiction of behaviorally relevant paths.
Cortical predictive coding models, in which separate populations encode predictions and prediction errors, also take on new significance when recast in time-symmetric terms. Prediction units send top-down signals that implement generative priors, while error units send bottom-up signals that encode deviations from those priors. Anatomically, these populations are embedded within recurrent loops that involve superficial and deep layers, feedback and feedforward pathways, and local inhibitory motifs. If prediction error units integrate information about mismatches that may be revealed only later in a sequence, then their activity contains backward-looking constraints. Plasticity that couples prediction and error units bilaterally can thus implement synaptic changes proportional to the overlap between forward-predicted and backward-corrected components. This arrangement remains within known cortical microcircuit constraints while realizing a form of credit assignment that depends on entire sequences of mismatches, not just instantaneous discrepancies.
Motor control circuits provide another domain in which time-symmetric learning offers a plausible explanation for observed physiology. The cerebellum, basal ganglia, and motor cortex are all heavily recurrent and receive delayed sensory and reward feedback about the consequences of actions. Yet organisms can rapidly adapt their motor commands to correct systematic errors, such as prism-induced visual displacements or force-field perturbations. A time-symmetric account posits that internal forward models within these circuits generate predictions of sensory outcomes, while delayed errors propagate back through recurrent loops to adjust the synapses that contributed to earlier parts of the command trajectory. Eligibility traces in parallel fiber–Purkinje synapses, dopaminergic modulation in basal ganglia pathways, and dendritic integration in motor cortical neurons together implement a distributed approximation to trajectory-level gradient descent. Crucially, no explicit storage of entire action histories is required; the combination of recurrent dynamics and synaptic memory at multiple timescales suffices to implement learning that is sensitive to both the past and eventual outcomes.
Biophysical constraints such as conduction delays, noise, metabolic costs, and limited precision of synaptic changes inevitably break exact time-reversal invariance in biological systems. However, time-symmetric principles do not require perfect symmetry; they require only that the effective equations governing neural trajectories and plasticity be approximately invariant under exchanging past and future, at the level of a coarse-grained description. From this standpoint, features like balanced excitation–inhibition, redundant parallel pathways, and normalization mechanisms can be interpreted as compensatory structures that keep the dynamics close to a symmetric regime. Synaptic scaling, homeostatic plasticity, and structural remodeling prune and reinforce connections in a way that preserves the capacity of recurrent circuits to propagate both predictive and corrective information without becoming unstable or dominated by noise.
Developmental processes can further bias circuits toward time-symmetric operation. Early spontaneous activity patterns and unsupervised exposure to environmental statistics help shape priors embedded in the connectivity of sensory and association cortices. Later, task-specific feedback, rewards, and social signals impose additional boundary conditions that refine these priors into task-appropriate internal models. If learning during development is guided by plasticity rules that integrate over entire episodes or behavioral bouts, then synapses are selected for their ability to support recurrent trajectories that remain robust under both forward-driven exploration and backward-driven correction. This developmental tuning may help explain why adult cortical networks can rapidly integrate new information into existing knowledge without catastrophic interference: their connectivity has been sculpted to encode structure that is stable under ongoing bidirectional refinement.
These considerations also bear on the longstanding debate about the plausibility of gradient-based learning in the brain. Classical objections emphasize the difficulty of transmitting precise error derivatives backward through fixed weights and of matching feedback weights to feedforward ones. Time-symmetric frameworks sidestep many of these issues by positing that feedback and feedforward interactions arise from the same pool of recurrent connections and that learning depends on locally available correlations shaped by the natural relaxation dynamics of the network. Feedback does not need to carry exact gradients; it needs only to carry constraint information that, when combined with stored eligibility traces, drives synaptic changes in roughly the right direction in trajectory space. Redundancy in connectivity, distributed representations, and slow synaptic consolidation can then average out noise and approximation error, yielding effective gradient-like behavior over longer timescales even if individual updates are coarse.
Experimental tests of time-symmetric learning hypotheses will likely require interventions that selectively perturb either forward- or backward-like components of neural activity. For example, artificially disrupting late-arriving feedback during sequence tasks while leaving early sensory drive intact should impair the ability of circuits to incorporate future constraints into present representations, reducing performance on tasks that require smoothing rather than simple prediction. Conversely, manipulating dendritic integration or neuromodulatory timing to decorrelate eligibility traces from delayed evaluative signals should degrade trajectory-level learning while leaving instantaneous responses relatively unchanged. Recordings that track the evolution of population activity across entire trials, combined with causal perturbations of specific circuit motifs, can reveal whether internal trajectories are being reshaped in ways consistent with bidirectional constraint propagation.
Many anatomical and physiological features of real neural systems—rich recurrent connectivity, layered feedforward–feedback loops, diverse inhibitory interneurons, dendritic compartmentalization, neuromodulatory control, oscillations, and replay—can be reinterpreted as components of an approximate time-symmetric learning machine. Rather than implementing separate mechanisms for prediction and error backpropagation, biological circuits appear well positioned to embed both roles in their ongoing dynamics and plasticity. While the brain does not compute exact gradients, it may exploit these structural motifs to approximate trajectory-level optimization under priors and outcomes, achieving functionally gradient-like adaptation through mechanisms that are entirely local, noisy, and constrained by biophysics, yet remarkably effective for real-world tasks.
