Dual Propagation: Accelerating Contrastive Hebbian Learning with Dyadic Neurons

Authors: Rasmus Høier, D. Staudt, Christopher Zach

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimentally we show on common computer vision datasets, including Imagenet32x32, that dual propagation performs equivalently to back-propagation both in terms of accuracy and runtime.
Researcher Affiliation Academia 1Chalmers University of Technology, Sweden. Correspondence to: Rasmus Høier <hier@chalmers.se>.
Pseudocode Yes Algorithm 1 Dual propagation
Open Source Code Yes Our code is available on github2. 2https://github.com/Rasmuskh/Dual-Propagation
Open Datasets Yes We evaluate dual propagation on MNIST, CIFAR10, CIFAR100, and Imagenet32x32. [...] The models were trained using standard data augmentation techniques (random-crops and horizontal flips) and the training data of 50 000 images was split into 45 000 for training and 5 000 for validation. [...] As Imagenet32x32 does not have a public test dataset we used the validation data as test data and reserved 5% of the training data for validation.
Dataset Splits Yes 10% of the training data was reserved for validation, and performance on the validation data was used to select which checkpoint of the model to evaluate on the test dataset. [...] the training data of 50 000 images was split into 45 000 for training and 5 000 for validation. [...] reserved 5% of the training data for validation.
Hardware Specification Yes On an NVIDIA A100 GPU both BP and DP had runtimes of 3.5 seconds per epoch and 4.5 seconds per epoch for CIFAR10 and CIFAR100 respectively [...] The experiments were enabled by the supercomputing resource Berzelius provided by National Supercomputer Centre at Link oping University and the Knut and Alice Wallenberg foundation.
Software Dependencies No The paper mentions 'autodiff framework' but does not specify software names with version numbers needed for replication.
Experiment Setup Yes The hyper-parameters used in the MLP experiments of Section 5.1 are listed in Table 5. [...] The hyper-parameters used in the experiments of Section 5.2 are listed in Table 6.