Towards the Fundamental Limits of Knowledge Transfer over Finite Domains
Authors: Qingyue Zhao, Banghua Zhu
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct simulations to verify the intuitive performance rankings bπCE,sgl bπSEL,ptl bπCEful given moderately large sample sizes and also numerically provide the asymptotical biasedness of bπCE,ptl with a finite-sample counterpart. |
| Researcher Affiliation | Academia | Qingyue Zhao Department of Computer Science and Technology Tsinghua University zhaotsingyue@gmail.com Banghua Zhu Department of Electrical Engineering and Computer Sciences University of California, Berkeley banghua@berkeley.edu |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. Methods are described textually and mathematically. |
| Open Source Code | No | The paper does not provide a specific link to open-source code for the methodology described. It cites other open-source models but does not offer its own implementation code. |
| Open Datasets | No | The paper describes simulated data generation using "Instance 0", "Instance 1", "Instance 2", and "Instance 3" (e.g., "Instance 0 For every s S, π ( |s) := 0.5Uniform(A)+0.5Dirac(A,s mod A+1)"). These are theoretical data generating processes for simulations, not publicly available datasets with access information. |
| Dataset Splits | No | The paper does not provide specific details about training, validation, or test dataset splits. It discusses sample sizes (n) for simulations based on theoretical data distributions, rather than partitioning an external dataset. |
| Hardware Specification | No | The paper does not provide any specific hardware details such as CPU, GPU models, or cloud computing specifications used for running the experiments. |
| Software Dependencies | No | The paper does not list specific software dependencies with version numbers that would be necessary for reproduction. |
| Experiment Setup | Yes | In this section, we specify a fair inductive bias due to the tabular nature: if s / S(D), bπ( |s) is set to Uniform(A) for all learners; for bπSEL,ptl( |s), the missing mass is amortized uniformly among A\A(D,s) if s S(D). Each marker in Figure 1 represents the empirical mean of TV(bπ,π |ρ) in 100 independent repeats given the corresponding sample size n. In Figure 2, in which each marker represents the average of 64000 independent repeats. |