Deep Learning from Crowds

Authors: Filipe Rodrigues, Francisco Pereira

AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically show that the proposed approach is able to internally capture the reliability and biases of different annotators and achieve new state-of-the-art results for various crowdsourced datasets across different settings, namely classification, regression and sequence labeling.
Researcher Affiliation Academia Filipe Rodrigues, Francisco C. Pereira Dept. of Management Engineering, Technical University of Denmark Bygning 116B, 2800 Kgs. Lyngby, Denmark rodr@dtu.dk, camara@dtu.dk
Pseudocode No The paper describes algorithms in text, but no structured pseudocode or algorithm blocks were found.
Open Source Code Yes Source code, datasets and demos for all experiments are provided at: http://www.fprodrigues.com/.
Open Datasets Yes Source code, datasets and demos for all experiments are provided at: http://www.fprodrigues.com/. (Kaggle 2013); (Rodrigues et al. 2017); (Russell et al. 2008); (Rodrigues, Pereira, and Ribeiro 2013); (Sang and Meulder 2003)
Dataset Splits Yes The base architecture was selected from a set of possible configurations using the true labels by optimizing the accuracy on a validation set (consisting of 20% of the train set) through random search.
Hardware Specification Yes reducing the training time by at least one order of magnitude when compared to the latter (minutes instead of several hours on a Core i7 with 32GB of RAM and a NVIDIA GTX 1070).
Software Dependencies No The proposed crowd layer (CL) was implemented as a new type of layer in Keras (Chollet 2015), so that using it in practice requires only a single line of code.
Experiment Setup Yes For this particular problem we used a fairly standard CNN architecture with 4 convolutional layers with 3x3 patches, 2x2 max pooling and Re LU activations. The output of the convolutional layers is then fed to a fully-connected (FC) layer with 128 Re LU units and finally goes to an output layer with a softmax activation. We use batch normalization (Ioffe and Szegedy 2015) and apply 50% dropout between the FC and output layers.