SaFARi: State-Space Models for Frame-Agnostic Representation
Authors: Hossein Babaei, Mel White, Sina Alemohammad, Richard Baraniuk
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate that Sa FARi can generate SSMs for function approximation over any frame or basis by choosing examples that are non-orthogonal, incomplete, or redundant. We then evaluate Sa FARi-generated state-space models on some sample datasets online function approximation, benchmarking against established baselines. |
| Researcher Affiliation | Academia | Hossein Babaei EMAIL Department of Electrical and Computer Engineering Rice University; Mel White EMAIL Department of Electrical and Computer Engineering Rice University; Sina Alemohammad EMAIL Department of Electrical and Computer Engineering Rice University; Richard G. Baraniuk EMAIL Department of Electrical and Computer Engineering Rice University |
| Pseudocode | No | The paper includes mathematical formulations and derivations but does not contain any clearly labeled pseudocode or algorithm blocks with structured, code-like steps. |
| Open Source Code | Yes | Code to replicate the results of this section, as well as generate SSMs with arbitrary frames is provided at: https://github.com/echbaba/safari-ssm. |
| Open Datasets | Yes | S&P 500: We use the daily S&P 500 index as a broad, large-cap U.S. equities benchmark over the last decade: from August 2015 to August 2025 (Yahoo Finance (2025)). The series consists of end-of-day levels for the price index. |
| Dataset Splits | No | The paper describes data preparation strategies like collecting 'overlapping sequences of 500 samples' and 'resampled into 4,000 samples' for the S&P 500 dataset, and setting a 'window size...at 10% of the input signal length' for the translated case. However, it does not explicitly provide traditional training, validation, and test dataset splits with specific percentages or counts. |
| Hardware Specification | No | The paper mentions general categories of hardware like 'parallel hardware such as GPUs' in discussions about computational complexity, but it does not specify any exact GPU/CPU models, processor types, or memory amounts used for running the experiments. |
| Software Dependencies | No | The paper mentions 'generalized bilinear transform (GBT)' and 'Adam optimizer' but does not provide specific version numbers for these or any other software components used in the experiments. |
| Experiment Setup | Yes | Both scaled and translated versions were evaluated with N = 32, 64, 128, where N is the size of the signal representation. For the translated case, the window size is set at 10% of the input signal length. The LSTM and GRU models are trained using an Adam optimizer (Kingma (2014)) until they converge, and the final validation performance is shown in Fig. 9. |