A Constraint-Based Approach to Learning and Explanation

Authors: Gabriele Ciravegna, Francesco Giannini, Stefano Melacci, Marco Maggini, Marco Gori3658-3665

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental An experimental evaluation is provided to support the proposed approach, in which we also explore the regularization effects introduced by the proposed Information Based Learning of Constraint (IBLC) algorithm.
Researcher Affiliation Academia 1Department of Information Engineering (DIINFO), University of Florence, Florence, Italy gabriele.ciravegna@unifi.it 2Department of Information Engineering and Science (DIISM), University of Siena, Siena, Italy {fgiannini, mela, maggini, marco}@diism.unisi.it
Pseudocode Yes Algorithm 1 Stage-based Optimization procedure (superscripts indicate the iteration number).
Open Source Code Yes All experiments are reproducible by using the code available at https://github.com/gabrieleciravegna/Information-based-Constrained-Learning
Open Datasets Yes The datasets can be downloaded at http://yann.lecun.com/ exdb/mnist/ and https://www.cs.stanford.edu/ roozbeh/pascal-parts/pascal-parts.html
Dataset Splits Yes Training, validation and test sets are composed of 20k, 5k, and 10k digits, respectively, taken from a (class-balanced) subset of the MNIST data.
Hardware Specification No The paper does not provide any specific details about the hardware used for the experiments, such as GPU or CPU models.
Software Dependencies No The paper mentions the 'Adam optimizer' but does not specify version numbers for it or any other software libraries, frameworks, or programming languages used.
Experiment Setup Yes The model parameters have been cross-validated ranging them in the following grids: learning rate of the Adam optimizer {0.01, 0.001, 0.0001}, γψ {1e 2, 1e 3, 1e 4}, γf = 1e 6 (we used the squared norm of the weights to implement f ). We also considered different scaling factors {0.5, 1, 2} of the penalty term associated to the MI (Eq. 4), and a convex combination of the two entropy terms, modulated by a coefficient {0.25, 0.5, 0.75}. All experiments are based on 1000 epochs. For the stage-based optimization, the task function learning stage lasts 200 epochs, while the constraint learning stage lasts 50 epochs.