Temporal Variational Implicit Neural Representations
Batuhan Koyuncu, Rachael DeVries, Ole Winther, Isabel Valera
TL;DR
This paper presents TV-INRs, a probabilistic neural framework for modeling and predicting irregular multivariate time series efficiently, especially in low-data scenarios, without extensive training or fine-tuning.
Contribution
Introduction of TV-INRs, a novel implicit neural representation framework that models distributions over time-continuous functions for improved imputation and forecasting.
Findings
Accurately performs diverse imputation and forecasting tasks with a single forward pass.
Outperforms existing methods by an order of magnitude in low-data regimes.
Offers a scalable and computationally efficient solution for real-world time series applications.
Abstract
We introduce Temporal Variational Implicit Neural Representations (TV-INRs), a probabilistic framework for modeling irregular multivariate time series that enables efficient individualized imputation and forecasting. By integrating implicit neural representations with latent variable models, TV-INRs learn distributions over time-continuous generator functions conditioned on signal-specific covariates. Unlike existing approaches that require extensive training, fine-tuning or meta-learning, our method achieves accurate individualized predictions through a single forward pass. Our experiments demonstrate that with a single TV-INRs instance, we can accurately solve diverse imputation and forecasting tasks, offering a computationally efficient and scalable solution for real-world applications. TV-INRs excel especially in low-data regimes, where it outperforms existing methods by an order of…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
* Well-founded probabilistic framework that rigorously combines VAE principles, amortized inference, and INRs * The model demonstrates superior results on sparse observations * The VAE implementation with conditional priors is technically sound, and the integration of Transformer encoders with hypernetwork-generated INRs is properly implemented * The covariates variant demonstrates that incorporating static covariates can improve performance at extreme sparsity thought gains are inconsisten
* The presented work is architecturally very similar to HyperTime (Fons et al., 2022), which combines set encoders, hypernetworks, and INRs for time series. I think the main differences in TV-INRs are: (1) using a Transformer encoder instead of DeepSets, and (2) adopting an explicit VAE framework with prior/posterior distributions instead of deterministic latent representations. There is very little discussion of how this work positions itself relative to HyperTime, which is detrimental to unde
- This paper is well written and easy to follow. Especially, clear notations make me easy to read this paper. - This paper includes extensive amounts of experiments in various experimental settings and clearly states the experimental details in the appendix. - It is interesting to see the extension of INR for time series to probabilistic modeling.
- **Limited novelty**: The idea for learning continuous function is very nice for learning irregularly sampled time-series data so many researchers have been adapted this method for time-series modeling as the author mentioned in related work section. And most of techniques employed in this work is not new. There is very few technical contribution except learning "Probabilistic model" using INR. I did not find any novel engineering in loss function, neural network architecture, preprocessing an
1) The idea of including the information of temporal stamps, feature vectors, and static covariates seems reasonable. 2) The proposed method is tested on different time series datasets and compared against different benchmarks. 3) Relevant ablation study results are given in the paper.
1) The so-called static covariates seem to provide conditional important information for the generative model. But the essential technical details regarding the static covariates for different data types are missing in the paper. 2) The authors proposed to use the temporal stamps information to model the sequential dependency, which is similar to Neural ODEs and diffusion models. Is this approach more suitable for modeling continuous systems than discrete systems? How is it compared to the posi
1. The choice to use Implicit Neural Representations (INRs) for time series modeling — including imputation, forecasting, and handling irregular sampling — is elegant. The paper is generally well-written, and the appendices are detailed and comprehensive. 2. Compared to prior INR-based approaches for time series, the proposed model is naturally suited to handle covariates and multivariate data, which is a clear improvement. 3. The ability to handle varying levels of missingness (in imputation ta
1. The paper’s first main claim is the introduction of a probabilistic framework for INR-based time series modeling, but this aspect is neither clearly motivated nor actually leveraged in the experiments. It could have been justified through, for example, quantile prediction and uncertainty estimation metrics through the WQL metric. 2. The second claim concerns inference “without requiring per-sample optimization,” in contrast to meta-learning methods like TimeFlow or DeepTime. While TV-INR achi
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsNeural Networks and Applications