Embedded Files
Content
Experiment data
Experiment data
Overall, our visualization technique can well
(1) preserve the high-dimensional neighbours in low dimensional space.
(2) preserve the high-dimensional distance between training points and classification boundary in low dimensional space.
(3) reconstruct a low-dimensional point into its original high-dimensional version, and
(4) preserve the visualization effect between consecutive training epochs.
In this section, we show the all experiment results as we stated in DVI paper.
NN preserving property
NN preserving property
our method is comparable with other neighbor preserving algorithms. Although t-SNE signigicantly outperforms other methods, it cannot genelize the projection to any unseen samples.
k=10
k=15
k=20
Boundary preserving property
Boundary preserving property
DVI significantly surpass other methods.
k=10
k=15
k=20
PPR
PPR
DVI vs DeepView
DVI vs DeepView
DVI outperform DeepView in terms of all properties and runtime efficiency.
k=10
k=15
k=20
Additional results on EMNIST dataset.
Additional results on EMNIST dataset.
We complement our experiments with one additional dataset suggested by reviewer#1.
Due to the time limit, we downsampled the size of EMNIST to be 50000 and conducted the experiments.
The quantitative results are available at link.
The qualitative results are shown below.
We show all 50000 training data here, each is represented by a small dot. Their colors represent label.
Those mispredicted samples are highlighted with a bigger circle, with inner colors being their prediction and outer colors being their true labels.
Observations:
The training process will push samples from different classes far away.
We observed that all classes stay in the middle at the beginning and gradually separate from each other, especially class "9" and "4".
The training process will pull samples from the same class together.
From epoch 1-3, the "6" (pink) class has two small clusters, but they merge at epoch 5. The same phenomenon can be observed for class "3"(red).
Most of the mispredicted samples lie on decision boundaries.
For example, at epoch 4, there are some mispredicted samples between class "4"(purple) and class "9"(cyan). They lie perfectly on the decision boundary between these two classes. This indicated our DVI's effectiveness on drawing decision boundaries.
Epoch 1
Epoch 2
Epoch 3
Epoch 4
Epoch 5
Epoch 6
Epoch 7
Epoch 8
Page updated
Google Sites
Report abuse