Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Recurrent Memory for Online Interdomain Gaussian Processes
Authors: Wenlong Chen, Naoki Kiyohara, Harrison Zhu, Jacob Curran-Sebastian, Samir Bhatt, Yingzhen Li
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate OHSVGP with online prediction for 1D time series, continual learning in discriminative GP model for data with multidimensional inputs, and deep generative modeling with sparse Gaussian process variational autoencoder, showing that it outperforms existing online GP methods in terms of predictive performance, long-term memory preservation, and computational efficiency. |
| Researcher Affiliation | Collaboration | Wenlong Chen1, , Naoki Kiyohara1,2, , Harrison Bo Hua Zhu3,1, , Jacob Curran-Sebastian3, Samir Bhatt3,1, Yingzhen Li1, 1Imperial College London 2Canon Inc. 3University of Copenhagen |
| Pseudocode | Yes | Algorithm 1 The HIPPO-SVGP ELBO for a single task of data. Differences with SVGP in blue. Require: X = {x1, . . . xn = t1} (training time steps up to time t1), {y1, . . . , yn} (training targets), Z RM 1 (inducing points) A(t) RM M, B(t) RM 1 (HIPPO matrices) , mu RM 1, Su RM M (variational params) 1: Kfu = k(X, Z), Kuu = k(Z, Z), Kt1 fu, Kt1 uu from HIPPO ODEs evolved from 0 to the final time step t1 with HIPPO matrices A(t), B(t) 2: µ(xi) = Kt1 fiu(Kt1 uu) 1mu Variational Posterior Mean 3: σ2(xi) = Kt1 fiu(Kt1 uu) 1[Kt1 uu Su](Kt1 uu) 1Kt1 ufi Variational Posterior Variance 4: ℓvarexp Pn i=1 EN(µ(xi),σ2(xi)) log p(yi | fi) closed form or quadrature/MC 5: KL KL(N(mu, Su)||N(0, Kt1 uu)) 6: return ℓvarexp KL |
| Open Source Code | Yes | Source Code: https://github.com/harrisonzhu508/HIPPOSVGP. ... We have released our code at https://github.com/harrisonzhu508/HIPPOSVGP/tree/main. |
| Open Datasets | Yes | Time series prediction. We consider regression benchmarks, Solar Irradiance [Lean, 2004], and Audio Signal [Bui and Turner, 2014] produced from the TIMIT database [Garifolo et al., 1993]. In addition, we consider a daily death-count time series from Santa Catarina State, Southern Brazil spanning the March 2020 to February 2021 COVID-19 pandemic, obtained from Hawryluk et al. [2021]. ... Continual learning. We consider continual learning on two UCI datasets with multi-dim inputs, Skillcraft [Blair et al., 2013] and Powerplant [Tfekci and Kaya, 2014]... High dimensional time series prediction. We evaluate GPVAEs on hourly climate data from ERA5 [Copernicus Climate Change Service, Climate Data Store, 2023, Hersbach et al., 2023] |
| Dataset Splits | Yes | We construct online learning tasks by splitting each dataset into 10 (5 for COVID) sequential partitions with an equal number of training instances. ... We construct two types of continual learning problems by first sorting the data points based on either the values in their first dimension or their L2 distance from the origin, and then splitting the sorted datasets into 10 sequential tasks with an equal number of training instances. ... The dataset is split into 10 sequential tasks of 186 hourly time steps each. |
| Hardware Specification | Yes | Table 1: Wall-clock accumulated runtime for learning all the tasks on a single NVIDIA RTX3090 GPU in seconds... on single NVIDIA A6000 GPU. |
| Software Dependencies | No | All the models are trained using Adam [Kingma and Ba, 2015] with the same learning rate and number of iterations. |
| Experiment Setup | Yes | Within each set of experiments, all the models are trained using Adam [Kingma and Ba, 2015] with the same learning rate and number of iterations. For OHSVGP, we construct inducing variables based on Hi PPO-Leg S [Gu et al., 2020]... and use 1000 RFF samples. We use ARD-RBF kernel, except for OVFF, tailored specifically to Mat ern kernels, where we use Mat ern5 2 kernel instead. ... All methods use M {15, 30} inducing points and are trained for 5000 iterations per task with a learning rate of 0.01. ... We use 256 inducing variables for all methods, and for each task, we train each method for 2000 iterations with a learning rate of 0.005. ... we use M {50, 100} and train each task for 20 epochs with learning rate 0.005 on single NVIDIA A6000 GPU. |