Fast and Flexible Multi-Task Classification using Conditional Neural Adaptive Processes

Authors: James Requeima, Jonathan Gordon, John Bronskill, Sebastian Nowozin, Richard E. Turner

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate that CNAPS achieves state-of-the-art results on the challenging META-DATASET benchmark indicating high-quality transfer-learning. We show that the approach is robust, avoiding both over-fitting in low-shot regimes and under-fitting in high-shot regimes. Timing experiments reveal that CNAPS is computationally efficient at test-time as it does not involve gradient based adaptation. Finally, we show that trained models are immediately deployable to continual learning and active learning where they can outperform existing approaches that do not leverage transfer learning.
Researcher Affiliation Collaboration James Requeima University of Cambridge Invenia Labs jrr41@cam.ac.uk Jonathan Gordon University of Cambridge jg801@cam.ac.uk John Bronskill University of Cambridge jfb54@cam.ac.uk Sebastian Nowozin Google Research Berlin nowozin@google.com Richard E. Turner University of Cambridge Microsoft Research ret26@cam.ac.uk
Pseudocode Yes Algorithm A.1 details computation of the stochastic estimator for a single task.
Open Source Code Yes Source code available at https://github.com/cambridge-mlg/cnaps.
Open Datasets Yes The first experiment tackles a demanding few-shot classification challenge called META-DATASET [6].
Dataset Splits Yes In our experiments we use the protocol defined by Triantafillou et al. [6] and META-DATASET for this sampling procedure. The TEXTURES dataset, which has only seven test classes and accuracy is highly sensitive to the train / validation / test class split.
Hardware Specification Yes The overall time required to produce the plot was 1274 and 7214 seconds for CNAPS and gradient approaches, respectively, on a NVIDIA Tesla P100-PCIE-16GB GPU.
Software Dependencies No The paper mentions general software concepts like 'gradient based adaptation' but does not specify particular software libraries or versions (e.g., PyTorch 1.9, TensorFlow 2.0).
Experiment Setup Yes The experiments use the following modelling choices (see Appendix E for full details). While CNAPS can utilize any feature extractor, a Res Net18 [14] is used throughout to enable fair comparison with Triantafillou et al. [6].