MNIST Multitask Save
6️⃣6️⃣6️⃣ Reproduce ICLR '18 under-reviewed paper "MULTI-TASK LEARNING ON MNIST IMAGE DATASETS"
Project README
MNIST-multitask
Reproduce ICLR2018 under-reviewed paper MULTI-TASK LEARNING ON MNIST IMAGE DATASETS
the paper argues that pre-train network with MNIST-like dataset can boost performance
results
| dataset | single-task | M+F | single-task (paper reported) | M+F (paper reported) |
|---|---|---|---|---|
| MNIST | 0.996 | 0.9956 | 0.9956 | 0.9971 |
| FashionMNIST | 0.9394 | 0.942 | 0.9432 | 0.9518 |
discussion
in my reproduction, FashionMNIST performs better with MNIST+FashionMNIST pre-trained first
but MNIST doesn't enjoy the benefits of pre-training.
The bias between reproduction and paper can result from preprocess of data.
Author
Po-Chih Huang / @pochih
Open Source Agenda is not affiliated with "MNIST Multitask" Project. README Source: pochih/MNIST-multitask
Stars
42
Open Issues
2
Last Commit
5 years ago
Repository
Open Source Agenda Badge
