Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]

Enhancing deep neural networks through complex-valued representations and Kuramoto synchronization dynamics

Authors: Sabine Muzellec, Andrea Alamia, Thomas Serre, Rufin VanRullen

TMLR 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate two architectures employing synchrony: a feedforward model and a recurrent model with feedback connections to refine phase synchronization using top-down information. Both models outperform a real-valued baseline and complex-valued models without Kuramoto synchronization on tasks involving multiobject images, such as overlapping handwritten digits, noisy inputs, and out-of-distribution transformations. Our findings highlight the potential of synchrony-driven mechanisms to enhance deep learning models, improving their performance, robustness, and generalization in complex visual categorization tasks.
Researcher Affiliation Academia Sabine Muzellec EMAIL Cer Co CNRS, University of Toulouse, France Carney Institute for Brain Science, Brown University, USA Andrea Alamia EMAIL Cer Co CNRS, University of Toulouse, France Thomas Serre EMAIL Carney Institute for Brain Science, Brown University, USA Rufin Van Rullen EMAIL Cer Co CNRS, University of Toulouse, France
Pseudocode Yes A.1 Algorithms We detail here the different steps of both versions of Komplex Net using the pseudo-code algorithm. Algorithm 1 describes the operations of Komplex Net and Algorithm 2 specifies how feedback connections integrate in the previous dynamic. Algorithm 1 Komplex Net Algorithm 2 Komplex Net with feedback
Open Source Code Yes Code available at https://github.com/S4b1n3/Komplex Net
Open Datasets Yes The first one is the Multi-MNIST dataset (Sabour et al., 2017), consisting of images containing two hand-written digits taken from the MNIST dataset. ... Similarly, we generate a version of this dataset with greyscaled CIFAR10 (Krizhevsky et al., 2009) images in the background.
Dataset Splits No The paper mentions generating an overlapping and non-overlapping version of the Multi-MNIST dataset and using greyscaled CIFAR10 images in the background. It also states that models are 'trained with two or three non-overlapping digits' and evaluated on 'in-distribution images (two-digit, non-overlapping images, all different from the training set)'. It uses a 'validation set' to select the best model. However, specific split percentages or exact counts for training, validation, and test sets are not provided.
Hardware Specification Yes All experiments are implemented in Pytorch 1.13 (Paszke et al., 2017) and run on a single NVIDIA V100. ... We acknowledge the Cloud TPU hardware resources that Google made available via the Tensor Flow Research Cloud (TFRC) program as well as computing hardware supported by NIH Office of the Director grant S10OD025181.
Software Dependencies Yes All experiments are implemented in Pytorch 1.13 (Paszke et al., 2017)
Experiment Setup Yes We train each family of models end-to-end using Adam (Kingma & Ba, 2014), a fixed learning rate of 1e 3, and a batch size of 128 or 32 depending on the dataset. ... The concerned parameters are: the desynchronization term ϵ, the gain parameters ηl for each layer l, the coupling kernel sizes kl0 and kl1, and the balance of the losses τ. ... The Vi T was trained for 50 epochs longer than the 30 epochs used for other models to ensure convergence.