Imbedding Deep Neural Networks

Authors: Andrew Corbett, Dmitry Kangin

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

Reproducibility Variable Result LLM Response
Research Type Experimental Through a series of experiments, we show the competitive performance of the proposed architectures for supervised learning and time series prediction. Accompanying code is made available at github.com/andrw3000/inimnet. In this section we demonstrate the practical ability of the proposed architectures in solving benchmark problems for deep neural networks.
Researcher Affiliation Collaboration Andrew Corbett University of Exeter a.j.corbett@exeter.ac.uk Dmitry Kangin Etcembly Ltd. dkangin@gmail.com
Pseudocode Yes Algorithm 1 Independent forward and backward pass with In Im Net
Open Source Code Yes Accompanying code is made available at github.com/andrw3000/inimnet.
Open Datasets Yes In the Rotating MNIST experiment, we solve the task of learning to generate the handwritten digit 3 for a sequence of sixteen equally spaced rotation angles between [0, 2π] given only the first example of the sequence. As in the previous section, we replicate the experimental setting of Yıldız et al. (2019) and Vialard et al. (2020).
Dataset Splits No No explicit description of training/validation/test dataset splits (percentages or counts) was found within the paper's own experimental setup. It references experimental settings of prior work but doesn't detail the splits used in its own context.
Hardware Specification No We measured the time per epoch whilst using a public configuration of Google Colab for the (best-performing) up-down method of Vialard et al. (2020) against In Im Net (with pmin = 3; three-layer MLP).
Software Dependencies Yes We reproduce the experimental setting from Vialard et al. (2020); the experiments have been conducted in Tensorflow 2.6.0.
Experiment Setup Yes The hyperparameters of the rotating MNIST model are listed in Table 2. The hyperparameters of the bouncing balls model are given in Table 3. The hyperparameters for the described convolutional model are listed in Table 4.