Active Transfer Learning under Model Shift
Authors: Xuezhi Wang, Tzu-Kuo Huang, Jeff Schneider
ICML 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the algorithms on synthetic functions and a realworld task on estimating the yield of vineyards from images of the grapes. We evaluate our methods on synthetic data and real-world grape image data. The experimental results show that our transfer learning algorithms significantly outperform covariate-shift methods with few labeled target data points, and our combined active transfer learning algorithm transfers knowledge from the source data and makes target labeling requests that achieve better prediction performance on the target data than alternative methods. |
| Researcher Affiliation | Academia | Xuezhi Wang XUEZHIW@CS.CMU.EDU Computer Science Department, Carnegie Mellon University Tzu-Kuo Huang TZUKUOH@CS.CMU.EDU Robotics Institute, Carnegie Mellon University Jeff Schneider SCHNEIDE@CS.CMU.EDU Robotics Institute, Carnegie Mellon University |
| Pseudocode | Yes | Algorithm 1 Conditional Distribution Matching 1: Input: Xtr, Y tr, {Xte L, Y te L} Xte U 2: Initialize w = 1, b = 0 3: repeat 4: Predict ˆY te U using {Xtr, Y new} {Xte L, Y te L}, where Y new is transformed using current w, b 5: Optimize the objective function in Equation 1 6: until w, b converge 7: Output: Prediction ˆY te U |
| Open Source Code | Yes | Target/Conditional shift, proposed by (Zhang et al., 2013), code is from http://people.tuebingen.mpg.de/ kzhang/Code-Tar S.zip. |
| Open Datasets | No | We have two datasets with grape images taken from vineyards and the number of grapes on them as labels, one is riesling (128 labeled images), another is traminette (96 labeled images), as shown in Figure 10. The paper mentions these datasets but does not provide specific access information like a URL, DOI, or a formal citation that would lead to public availability. |
| Dataset Splits | Yes | On the traminette dataset we have achieved R-squared correlation 0.754 (95% for training and 5% for test). |
| Hardware Specification | No | The paper does not provide any specific hardware details such as GPU/CPU models, memory, or cloud instance types used for running experiments. |
| Software Dependencies | No | The paper mentions using 'matlab notation' for synthetic datasets and 'Random Kitchen Sinks' for feature extraction, but does not specify version numbers for any software or libraries. |
| Experiment Setup | No | The paper mentions 'Parameters (kernel width, regularization term, etc.) are set using cross validation' but does not provide the specific values of these hyperparameters or other training configurations in the main text. |