Wandering within a world: Online contextualized few-shot learning

Authors: Mengye Ren, Michael Louis Iuzzolino, Michael Curtis Mozer, Richard Zemel

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we show experimental results for our online contextualized few-shot learning paradigm, using Roaming Omniglot and Roaming Rooms (see Sec. 3) to evaluate our model, CPM, and other state-of-the-art methods.
Researcher Affiliation Collaboration Mengye Ren1,3 Michael L. Iuzzolino2 Michael C. Mozer2,4 Richard S. Zemel1,3,5 1University of Toronto 2Google Research 3Vector Institute 4University of Colorado, Boulder 5CIFAR
Pseudocode No The paper describes the model (Contextual Prototypical Memory Networks) using text and mathematical equations, but it does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Our code and dataset are released at: https://github.com/renmengye/oc-fewshot-public
Open Datasets Yes Third, we build three datasets: 1) Roaming Omniglot is based on handwritten characters from Omniglot (Lake et al., 2015); 2) Roaming Image Net is based on images from Image Net (Russakovsky et al., 2015); and 3) Roaming Rooms is our new few-shot learning dataset based on indoor imagery (Chang et al., 2017), which resembles the visual experience of a wandering agent.
Dataset Splits Yes We split the alphabets into 31 for training, 5 for validation, and 13 for testing.
Hardware Specification No The paper mentions training across multiple GPUs ('across 2 GPUs', 'across 4 GPUs') but does not specify the model or type of these GPUs, nor does it detail any other specific hardware components like CPU models, memory, or accelerator types.
Software Dependencies No The paper mentions using 'the Adam optimizer (Kingma & Ba, 2015)' but does not specify version numbers for any software dependencies, such as programming languages (e.g., Python), deep learning frameworks (e.g., PyTorch, TensorFlow), or specific library versions.
Experiment Setup Yes Implementation details: For Roaming Omniglot, we use the common 4-layer CNN for few-shot learning with 64 channels in each layer. For Roaming Image Net, we also use Res Net-12 with input resolution 84x84 (Oreshkin et al., 2018). For the Roaming Rooms, we resize the input to 120x160 and use Res Net-12. ... For the contextual RNN, in both experiments we used an LSTM (Hochreiter & Schmidhuber, 1997) with a 256d hidden state. The best CPM model is equipped using GAU and cosine similarity for querying prototypes. Logits based on cosine similarity are multiplied with a learned scalar initialized at 10.0 (Oreshkin et al., 2018). We use the Adam optimizer (Kingma & Ba, 2015) for all of our experiments, with a gradient cap of 5.0. For Roaming Omniglot we train the network for 40k steps with a batch size 32 and maximum sequence length 150 across 2 GPUs and an initial learning rate 2e-3 decayed by 0.1 at 20k and 30k steps. For Roaming Rooms we train for 20k steps with a batch size 8 and maximum sequence length 100 across 4 GPUs and an initial learning rate 1e-3 decayed by 0.1 at 8k and 16k steps. We use the BCE coefficient λ = 1 for all experiments.