Learning to Learn with Conditional Class Dependencies
Authors: Xiang Jiang, Mohammad Havaei, Farshid Varno, Gabriel Chartrand, Nicolas Chapados, Stan Matwin
ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We use mini Image Net to evaluate the proposed Conditional class-Aware Meta-Learning algorithm. mini Image Net (Vinyals et al., 2016) is composed of 84 84 colored images from 100 classes, with 600 examples in each class. We adopt the class split by Ravi & Larochelle (2016) that uses 64 classes for training, 16 for validation, and 20 for test. For N-way K-shot training, we randomly sample N classes from the meta-train classes each containing K examples for training and 20 examples for testing. At meta-testing time, we randomly sample 600 N-way K-shot tasks from the test classes. The results presented in Table 1 show that our proposed algorithm has comparable performance on the state-of-the-art mini Image Net 5-way 1-shot classification task, and competitive results on the 5-way 5-shot task. |
| Researcher Affiliation | Collaboration | Imagia Inc., Dalhousie University {xiang.jiang,mohammad,farshid.varno,gabriel,nic}@imagia.com, stan@cs.dal.ca |
| Pseudocode | No | The paper describes the proposed Conditional class-Aware Meta-Learning (CAML) and its components in detail within the text and using figures (e.g., Figure 1), but it does not provide any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | No | The paper does not contain any explicit statements about releasing source code for the methodology, nor does it provide a link to a code repository. |
| Open Datasets | Yes | We use mini Image Net to evaluate the proposed Conditional class-Aware Meta-Learning algorithm. mini Image Net (Vinyals et al., 2016) is composed of 84 84 colored images from 100 classes, with 600 examples in each class. |
| Dataset Splits | Yes | We adopt the class split by Ravi & Larochelle (2016) that uses 64 classes for training, 16 for validation, and 20 for test. |
| Hardware Specification | No | The paper describes the model architecture and training details (e.g., number of layers, optimizers, episodes) but does not specify any particular hardware used for running the experiments (e.g., specific GPU models, CPU types, or cloud instances). |
| Software Dependencies | No | The paper mentions using the 'Adam optimizer (Kingma & Ba, 2014)' but does not provide specific version numbers for any software components, libraries, or programming languages used in the implementation. |
| Experiment Setup | Yes | The base-learner (fθ) is composed of 4 layers of 3 3 convolutions with a 4 4 skip connections from the input to the final convolutional layer. ... Each convolutional layer has 30 channels and is followed by CBN, Re LU and 2 2 max-pooling operations. ... The metric space is pre-trained on the same meta-training dataset for 30,000 episodes and not updated while learning the base-learner. The meta-learner is trained for 50,000 episodes. ... For CBN functions (fc), we use 3 dense layers with 30 hidden units each. Every layer is followed by a Re LU except for the last layer where no activation is used. For the meta-learner, we use MAML with 1 gradient step for 1-shot learning and 5 gradient steps for 5-shot learning. We use the Adam (Kingma & Ba, 2014) optimizer and clip the L2 norm of gradients with an upper bound of 5. |