Conditional Meta-Learning of Linear Representations

Authors: Giulia Denevi, Massimiliano Pontil, Carlo Ciliberto

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5 Experiments We now present preliminary experiments in which we compare the proposed conditional meta-learning approach in Alg. 1 (cond.) with the unconditional counterpart (uncond.) and solving the tasks independently (ITL, namely, running the inner algorithm separately across the tasks with the constant linear representation = Id 2 Sd +).
Researcher Affiliation Collaboration Giulia Denevi Leonardo Labs (Italy) giulia.denevi.ext@leonardo.com Massimiliano Pontil Istituto Italiano di Tecnologia (Italy) & University College of London (UK) massimiliano.pontil@iit.it Carlo Ciliberto University College of London (UK) & Istituto Italiano di Tecnologia (Italy) c.ciliberto@ucl.ac.uk
Pseudocode Yes Algorithm 1 Meta-algorithm, SGD on Eq. (15)
Open Source Code Yes We attached our code in the Supplementary material.
Open Datasets Yes We also tested the performance of the methods on the Movielens-100k and Jester-1 real-world datasets... We considered two variants of the setting described in Ex. 1 with side information corresponding to the training datasets Ztr associated to each task. ... on the regression problem on the computer survey data from [23]
Dataset Splits Yes The hyper-parameter γ was chosen by (meta-)cross validation on separate Ttr, Tva and Tte respectively meta-train, -validation and -test sets.
Hardware Specification Yes The experiments were run on a server with an Intel Xeon E3-1505M v5 CPU, 64 GB of RAM, and an NVIDIA Quadro M2000M GPU.
Software Dependencies No The paper does not specify particular software dependencies with version numbers (e.g., 'PyTorch 1.9' or 'TensorFlow 2.x'). It only generally mentions implementation in Python in Appendix E without further details.
Experiment Setup Yes The hyper-parameter γ was chosen by (meta-)cross validation on separate Ttr, Tva and Tte respectively meta-train, -validation and -test sets. For all experiments, we set n = 80, ntr = 40, nte = 40. The within-task learning rate (for the online algorithm in Eq. (3)) was set to η = 0.05 and the number of iterations to N = 1000. The batch size for the online algorithm was set to b = 1.