When data streams are interrupted—whether by a sensor failure, a patient missing a clinical scan, or a moving object passing behind an obstacle—we are left with a temporal gap. Standard autoregressive models approach this problem sequentially, predicting the next state based solely on historical data. While effective for forecasting the future, this unidirectional approach is fundamentally suboptimal for filling in the past.
The core intuition behind bidirectional temporal modeling is simple: the future constrains the past. If we know both where a trajectory began before a gap and where it resumed after it, the space of possible intermediate states shrinks dramatically. By anchoring the interpolation process at both ends of a missing segment, models can eliminate the compounding error typical of forward-only predictions.
The Limitation of Forward-Only Forecasting
To understand why bidirectional modeling works, consider the task of tracking an agent through a period of occlusion. A forward-looking model updates its hidden state chronologically. Once the agent disappears, the model must rely on its own previous predictions to generate the next step. This creates a feedback loop where small errors accumulate exponentially, causing the predicted path to drift rapidly away from reality.
In contrast, retrospective sequence completion—often called imputation or inpainting—operates in an offline or batch context where subsequent observations are already recorded. Ignoring these future data points throws away critical evidence. Knowing the exact time and location where an agent reappears provides a mathematical boundary condition that grounds the entire missing interval.
The Power of Temporal Anchoring
By integrating both preceding and succeeding context, bidirectional frameworks transform an open-ended prediction problem into a constrained interpolation problem.
As shown by Daniel Hennes et al. in their work on tracking sports players off-screen (Multiagent off-screen behavior prediction in football), anchoring a trajectory at both the disappearance and reappearance points drastically reduces error accumulation. The future state acts as a target that pulls the interpolated path back into alignment, ensuring a smooth and physically plausible transition across the gap.
Unidirectional (Forward Only):
[Known Past] ---> (Drifting Predictions...) ---> ? [Unknown Future]
Bidirectional Anchoring:
[Known Past] ---> (Constrained Interpolation) <--- [Known Future]
This dual-constraint mechanism is valuable across diverse domains:
- Mobility and Trajectories: In human mobility modeling, Dongbo Xi et al. demonstrate in Modelling of Bi-Directional Spatio-Temporal Dependence and Users' Dynamic Preferences for Missing POI Check-In Identification that knowing a user's subsequent destination narrow downs the probable intermediate locations they could have visited during a missing interval.
- Signal Processing: For high-frequency data streams like wireless sensing or radar, processing signals in both directions allows models to interpolate through localized interference bursts by drawing clean context from both sides of the corruption.
Unifying Forward and Backward Perspectives
Implementing this intuition requires architectures that can ingest and synthesize information from both directions without introducing data leakage or contradictory predictions.
A foundational approach to this problem is introduced in BRITS: Bidirectional Recurrent Imputation for Time Series by Wei Cao et al.. Instead of treating forward and backward passes as isolated streams, the architecture enforces a consistency loss between them. The forward recurrent network tracks historical decay, while the backward network back-propagates future constraints.
By combining these two perspectives, the model ensures that every missing value is estimated using a unified temporal context. This design prevents the step-by-step divergence common in unidirectional models, allowing the network to handle highly irregular sampling and prolonged data gaps.
Extending to Generative Frameworks
While recurrent networks provide a strong baseline for deterministic smoothing, complex data distributions—such as 3D medical imaging or non-linear physical dynamics—benefit from generative architectures that can model the underlying uncertainty of the missing data.
In state-space modeling, M. Fraccaro et al. (A Disentangled Recognition and Nonlinear Dynamics Model for Unsupervised Learning) demonstrate that integrating bidirectional smoothing mechanisms into variational autoencoders allows future observations to directly refine past latent states. This combination of deep learning for recognition and classical smoothing for dynamics yields highly accurate reconstructions of missing video frames.
More recently, this bidirectional conditioning strategy has been successfully adapted to diffusion models. For instance, Tianli Tao et al. developed LoCI-DiffCom: Longitudinal Consistency-Informed Diffusion Model for 3D Infant Brain Image Completion to fill in missing structural MRI scans across a developmental timeline. By feeding both preceding and subsequent scans into the diffusion generation process, the model uses the two known points as stable mathematical anchors. This allows the system to accurately interpolate rapid, non-linear anatomical changes while maintaining strict longitudinal consistency.
Ultimately, whether implemented via recurrent networks, state-space models, or generative diffusion, the core mechanism remains identical: by treating time not just as a forward-flowing arrow but as a fully observed coordinate space, bidirectional interpolation leverages the future to reconstruct the past.
Key Foundational Work
Predictive encoding of contextual relationships for perceptual inference (2015) — contributed 2 concept elements to this direction.
Using Social Dynamics to Make Individual Predictions: Variational Inference with a Stochastic Kinetic Model (2016) — contributed 2 concept elements to this direction.