An Expressive and Self-Adaptive Dynamical System for Efficient Function Learning
Authors: Chuan Liu, Chunshu Wu, Ruibing Song, Ang Li, Ying Nian Wu, Tong Geng
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results across diverse domains demonstrate that EADS achieves higher accuracy than existing works, while offering orders-of-magnitude speedups over traditional neural network solutions on GPUs for both inference and training, showcasing its broader impact in overcoming computational bottlenecks across various fields. Section 4. Evaluation |
| Researcher Affiliation | Academia | 1University of Rochester, Rochester, NY, USA 2Pacific Northwest National Laboratory, Richland, WA, USA 3University of Washington, Seattle, WA, USA 4University of California Los Angeles, Los Angeles, CA, USA 5Rice University, Houston, TX, USA. |
| Pseudocode | No | The paper describes its methodology using mathematical equations and textual explanations of system dynamics and training processes. It does not contain any clearly labeled pseudocode or algorithm blocks with structured, code-like formatting. |
| Open Source Code | No | The paper does not contain any explicit statements, direct links to repositories, or mentions of supplementary materials indicating that the source code for the methodology described is openly available. |
| Open Datasets | Yes | For complex function learning in real-world problems, we evaluate the performance of EADS in spatial-temporal prediction tasks including six real-world datasets from four applications. (1) Traffic flow prediction with two datasets PEMS04 and PEM08 (Chen et al., 2001). (2) Air quality prediction including PM2.5 and PM10 (Kong et al., 2021). (3) Taxi demand prediction (NYC Taxi): predicting the hourly number of taxi trips (New York City Taxi and Limousine Commission, 2024). (4) Pandemic progression prediction (Texas COVID): predicting the daily number of new cases (Centers for Disease Control and Prevention, 2024). ... we adopt the GPT-2 small model (Wolf, 2019)... on the LAMBADA dataset (Paperno et al., 2016). |
| Dataset Splits | No | The paper mentions evaluating performance and training models (e.g., 'report the test MAE', 'construct a training dataset'), but it does not explicitly provide the specific percentages, sample counts, or detailed methodology for how the datasets were split into training, validation, and test sets for all experiments. |
| Hardware Specification | Yes | We conduct our experiments using an NVIDIA A100 40GB SXM GPU for non-dynamical system based baselines, measuring total training time, inference latency per sample, and accuracy. For dynamical system based approaches, we build upon the original hardware embodiment BRIM (Afoakwa et al., 2021), using a custom CUDA-accelerated Finite Element Analysis (FEA) simulator to assess the training time, inference latency, and accuracy. Since the dynamical system based baseline NPGL (Wu et al., 2024) only achieves inference on dynamical systems, its training time is still measured on an A100 GPU using its own offline training method. |
| Software Dependencies | No | The paper mentions using a 'custom CUDA-accelerated Finite Element Analysis (FEA) simulator' and refers to the 'GPT-2 small model', implying software dependencies like CUDA or a deep learning framework, but it does not provide specific version numbers for any software components. |
| Experiment Setup | Yes | The number of hidden nodes in EADS is set to 128, and baselines are implemented following the experimental setups detailed in their respective original papers. Specifically, we adopt the GPT-2 small model (Wolf, 2019), which consists of 12 transformer decoders, each containing a causal self-attention kernel. |