Levi-Heath's picture
Update README.md
f157756 verified
|
raw
history blame
1.66 kB
metadata
license: cc-by-nc-nd-4.0
language:
  - en
pipeline_tag: image-classification
tags:
  - ImageNet-1k
  - image classification
datasets:
  - ILSVRC/imagenet-1k
metrics:
  - accuracy

Description of the ImageNet-1k Featured Model

For a thorough description of our models trained on the ImageNet-1k dataset, please read our preprint, Towards Errorless Training ImageNet-1k, which is available at [ADD LINK to arXiv preprint]. In ../ImageNet-1k/MATLAB, we give parameters for 6 models, which are listed in the table below. Each model has the following architecture: 17x40x2=1360 FNNs, all with homogeneous architecture (900-256-25 or 900-256-77-25), working in parrallel to produce 1360 predictions which determine a final prediction using the majority voting protocol. We trained models using the following transformation of the 64x64 downsampled ImageNet-1k dataset:

  • downsampled images to 32x32, using the mean values of non-overlapping 2x2 grid cells and
  • trimmed off top row, bottom row, left-most column, and right-most column.

This transformed data results in 30x30 images, hence 900-dimensional input vectors.

Model Training Method FNN Architecture Accuracy (%)
Model_S_h1_m1 SGD 900-256-25 98.247
Model_S_h1_m2 SGD 900-256-25 98.299
Model_S_h2_m1 SGD 900-256-77-25 96.990
Model_T_h1_m1 SGD followed by GDT 900-256-25 98.289
Model_T_h1_m2 SGD followed by GDT 900-256-25 98.300
Model_T_h2_m1 SGD followed by GDT 900-256-77-25 97.770
*SGD = stochastic gradient descent
**GDT = gradient descent tunneling