Deep Multi-task Representation Learning: A Tensor Factorisation Approach

Authors: Yongxin Yang, Timothy M. Hospedales

ICLR 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4 EXPERIMENTS
Researcher Affiliation Academia Yongxin Yang, Timothy M. Hospedales Queen Mary, University of London {yongxin.yang, t.hospedales}@qmul.ac.uk
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Our method is implemented with Tensor Flow (Abadi et al., 2015). The code is released on Git Hub4. For DMTRL-Tucker, DMTRL-TT, and DMTRL-LAF, we need to assign the rank of each weight tensor. https://github.com/wOOL/DMTRL
Open Datasets Yes We use MNIST handwritten digits. The Adience Faces (Eidinger et al., 2014) is a large-scale face images dataset... We next consider the task of learning to recognise handwritten letters in multiple languages using the Omniglot (Lake et al., 2015) dataset.
Dataset Splits Yes We then use cross-validation to select among the three userdefined MTL architectures... Running 5-fold cross-validation on the train set to evaluate the architectures...
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments.
Software Dependencies No Our method is implemented with Tensor Flow (Abadi et al., 2015).
Experiment Setup Yes We use a modified Le Net (Le Cun et al., 1998) as the CNN architecture. The first convolutional layer has 32 filters of size 5 5, followed by 2 2 max pooling. The second convolutional layer has 64 filters of size 4 4, and again a 2 2 max pooling. After these two convolutional layers, two fully connected layers with 512 and 1 output(s) are placed sequentially. The convolutional and first FC layer use RELU f(x) = max(x, 0) activation function. We use hinge loss, ℓ(y) = max(0, 1 ˆy y)...