Instance-Conditional Timescales of Decay for Non-Stationary Learning
Authors: Nishant Jain, Pradeep Shenoy
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on a large real-world dataset of 39M photos over a 9 year period show upto 15% relative gains in accuracy compared to other robust learning baselines. We replicate our gains on two collections of real-world datasets for nonstationary learning, and extend our work to continual learning settings where, too, we beat SOTA methods by large margins. |
| Researcher Affiliation | Industry | Google Research India {nishantjn, shenoypradeep}@google.com |
| Pseudocode | No | The paper describes mathematical formulations and optimization processes but does not include any explicitly labeled "Pseudocode" or "Algorithm" blocks. |
| Open Source Code | No | The paper refers to an appendix (Jain and Shenoy 2023, which is the arXiv version of this paper) for additional details, but does not explicitly state that code is released or provide a link to a code repository for the methodology described. |
| Open Datasets | Yes | We experiment extensively with the geolocalization (GEO) dataset proposed by (Cai, Sener, and Koltun 2021): 39M images from YFCC100M (Thomee et al. 2016)... Wild-Time Benchmark. This collection of 5 datasets captures real-world concept drift (Yao et al. 2022) Yearbook, FMo W-time, MIMIC, Huffpost, and arxiv... CLEAR Benchmark. We also experiment with the CLEAR-100 Benchmark (Lin et al. 2021)... |
| Dataset Splits | Yes | We split each dataset into train, validation, and test sets in temporal order, reusing existing partitions where available. Thus, validation data is more recent than train data. For the GEO dataset (see Fig. 1(a)), these sets contain 18M, 2M, 19M images covering time periods of around 54,6,60 months respectively. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used to run the experiments (e.g., GPU models, CPU types, memory specifications). |
| Software Dependencies | No | The paper does not list specific software dependencies with their version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | We make an additional design choice of setting ak = ak 0 for some fixed a0... (K = 16, a0 = 2 in our experiments)... All algorithms iterate over each bucket of data k times (k=5 for GEO, and as per the benchmark proposal for CLEAR100). |