/dcnet

Effective Abstract Reasoning with Dual-Contrast Network

Primary LanguagePython

Effective Abstract Reasoning with Dual-Contrast Network

Tao Zhuo, Mohan Kankanhalli

This code is the implementation of our ICLR 2021 paper

Results

Average testing accuracy of different models on RAVEN and neutral regime of PGM dataset. Aux means auxiliary annotations.

Method Aux Avg RAVEN PGM
ResNet-18+DRT - 59.56 -
WReN+Aux 55.44 33.97 76.90
LEN+Aux 70.85 59.40 82.30
MXGNet+Aux - - 89.60
ACL - 93.71 -
LSTM 24.44 13.07 35.80
CNN 34.99 36.97 33.00
WReN 40.10 17.94 62.60
Wild-ResNet - - 48.00
ResNet-50 64.13 86.26 42.00
MLRN 55.33 12.50 98.03
LEN 70.50 72.90 68.10
CoPINet 73.90 91.42 56.37
MXGNet 75.31 83.91 66.70
DCNet-RC 78.10 92.74 63.45
DCNet-CC 47.10 36.47 57.76
DCNet 81.08 93.58 68.57

Citation

@inproceedings{zhuo2021,
author={Tao Zhuo and Mohan Kankanhalli},
title={Effective Abstract Reasoning with Dual-Contrast Network},
booktitle={International Conference on Learning Representations (ICLR)},
year={2021} }