Memory-Associated Differential Learning

Memory-Associated Differential Learning (MAD Learning) learns the differential function among data rather than the mapping from inputs to outputs.

It is a flexible learning paradigm that can be applied to unary situations (Image Recognition), binary situations (Link Prediction) and ternary situations (Knowledge Graphs).

Currently, MAD Learning achieves competitive performance on obgl-ddi.

Reproducibility

Weekday Prediction Challenge

Run the following command to experiment MAD Learning in Weekday Prediction Challenge

python3 weekday.py

You may edit the weekday.py file from line 116 to line 126 to modify the hyperparameters of MAD Learning or conduct this experiment with MLP.

Image Recognition

python3 cifar_mnist.py cifar10 conv

The third and fourth parameters are dataset and method.

• available datasets: cifar10 / cifar100 / mnist / kmnist
• available methods: conv / resnet / mad-conv /mad-resnet

Zachary's Karate Club

To generate a series of points representing Zachary's Karate Club, run

python3 karate.py

Link Prediction

To predict links on ogbl-ddi, run python3 ogbl-ddi.py.

To predict links on Cora, Citeseer or Pubmed, edit file citations.py to set the dataset name at line 11 and run it.

Citation

@article{2102.05246v1,
  author = {Yi Luo and Aiguo Chen and Bei Hui and Ke Yan},
  eprint = {2102.05246v1},
  month = {Feb},
  title = {Memory-Associated Differential Learning},
  type = {article},
  url = {http://arxiv.org/abs/2102.05246v1},
  year = {2021},
}