Researchers have developed ICICLE, a new in-context surrogate model for curve-learning extrapolation in Bayesian optimization. ICICLE, with 1.28 million parameters, demonstrates comparable or superior performance to existing methods like FT-PFN and Quick-Tune on benchmarks such as LCBench, PD1, TaskSet, and Meta-Album. Notably, ICICLE achieves this without pre-training or meta-training on the Meta-Album benchmark, unlike Quick-Tune, and offers a more efficient approach by generating synthetic training tasks on the fly. AI
IMPACT This research could lead to more efficient hyperparameter optimization in machine learning workflows.
RANK_REASON The cluster contains a research paper detailing a new model and its performance on benchmarks. [lever_c_demoted from research: ic=1 ai=1.0]
- Bayesian optimization
- FT-PFN
- ICICLE
- Jax
- Laurits Fredsgaard
- LCBench
- LC-PFN
- Meta-Album
- ONNX
- PyTorch
- Quick-Tune
- Ray Tune
- TaskSet
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →