LEARNING TO PROPAGATE LABELS: TRANSDUCTIVE PROPAGATION NETWORK FOR FEW-SHOT LEARNING
Authors: Yanbin Liu, Juho Lee, Minseop Park, Saehoon Kim, Eunho Yang, Sung Ju Hwang, Yi Yang
ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate TPN on multiple benchmark datasets, on which it largely outperforms existing few-shot learning approaches and achieves the state-of-the-art results. |
| Researcher Affiliation | Collaboration | Yanbin Liu1 , Juho Lee2,3, Minseop Park3, Saehoon Kim3, Eunho Yang3,4, Sung Ju Hwang3,4 & Yi Yang1,5 1CAI, University of Technology Sydney, 2University of Oxford 3AITRICS, 4KAIST, 5Baidu Research |
| Pseudocode | No | The paper describes the steps of its approach in prose and through mathematical equations, but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not include any explicit statements about making its source code publicly available or provide a link to a code repository. |
| Open Datasets | Yes | We evaluate and compare our TPN with state-of-the-art approaches on two datasets, i.e., mini Image Net (Ravi and Larochelle, 2017) and tiered Image Net (Ren et al., 2018). The former is the most popular few-shot learning benchmark and the latter is a much larger dataset released recently for few-shot learning. |
| Dataset Splits | Yes | mini Image Net. ...It is composed of 100 classes randomly selected from Imagenet with each class containing 600 examples. In order to directly compare with state-of-the-art algorithms for few-shot learning, we rely on the class splits used by Ravi and Larochelle (2017), which includes 64 classes for training, 16 for validation, and 20 for test. All images are resized to 84 84 pixels. tiered Image Net. ...The top hierarchy has 34 categories, which are divided into 20 training (351 classes), 6 validation (97 classes) and 8 test (160 classes) categories. |
| Hardware Specification | No | The paper does not provide specific hardware details such as exact GPU/CPU models, processor types, or memory amounts used for running its experiments. It only mentions general network architectures and optimization methods. |
| Software Dependencies | No | The paper mentions using Adam optimizer but does not specify version numbers for any software libraries, frameworks, or programming languages used in the implementation of the experiments. |
| Experiment Setup | Yes | The hyper-parameter k of k-nearest neighbour graph (Section 3.2.2) is set to 20 and α of label propagation is set to 0.99, as suggested in Zhou et al. (2004). Following Snell et al. (2017), we adopt the episodic training procedure, i.e, we sample a set of N-way K-shot training tasks to mimic the N-way K-shot test problems. ... For 1-shot and 5-shot test problem, we adopt 5-shot and 10-shot training respectively. In all settings, the query number is set to 15 and the performance are averaged over 600 randomly generated episodes from the test set. All our models were trained with Adam (Kingma and Ba, 2015) and an initial learning rate of 10 3. For mini Image Net, we cut the learning rate in half every 10, 000 episodes and for tiered Image Net, we cut the learning rate every 25, 000 episodes. The reason for larger decay step is that tiered Image Net has more classes and more examples in each class which needs larger training iterations. We ran the training process until the validation loss reached a plateau. |