Multiclass Neural Network Minimization via Tropical Newton Polytope Approximation

Authors: Georgios Smyrnis, Petros Maragos

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

Reproducibility Variable Result LLM Response
Research Type Experimental In addition, we make experimental evaluations on the MNIST and Fashion-MNIST datasets, with our results demonstrating a significant reduction in network size, while retaining adequate performance.
Researcher Affiliation Academia School of ECE, National Technical University of Athens, Athens, Attiki, Greece. Also: Robot Perception and Interaction Unit, Athena Research Center, Maroussi, Greece.
Pseudocode Yes Algorithm 1 Heuristic Minimization Algorithm (Smyrnis et al., 2020), Algorithm 2 One-Versus-All Multiclass Minimization, Algorithm 3 Stable Divisor Picking Algorithm
Open Source Code Yes The code used is provided as supplementary material, and can also be found here: https://github.com/Georgios Smyrnis/ multiclass_minimization_icml2020.
Open Datasets Yes We made use of the MNIST dataset of handwritten digits (Lecun et al., 1998), as well as Fashion-MNIST (Xiao et al., 2017), to evaluate our methods (available at http://yann.lecun.com/exdb/mnist/ and https://github.com/zalandoresearch/ fashion-mnist, respectively).
Dataset Splits No The paper mentions training and testing but does not explicitly provide details about a validation dataset split.
Hardware Specification Yes The experiments were performed on a Core i5-7200U CPU clocked at 2.5GHz, with 8GB of RAM and 8GB of swap space.
Software Dependencies No For our experiments we made use of Keras for writing our code (Chollet et al., 2015). No specific version number for Keras or other software dependencies is provided.
Experiment Setup Yes For the MNIST dataset, the network consisted of two convolutional layers, each with 16 units, 5 5 kernels, Re LU activations and maxpooling with a factor of 3. The output of these layers was then fed to a two-layer feedforward network, with Re LU activations and 500 neurons in the hidden layer. For the Fashion-MNIST dataset, the structure was similar, except the convolutional layers contained 32 units and the hidden layer in the end 1000 neurons.