AutoEncoder for Fashion-MNIST

• 

  Training and Validation Loss over Epochs

Inspiration

Having worked in the field of Machine Learning for some time, we realized that labeled data was hard and expensive. In order to overcome this problem, we started searching for ideas that would help optimize the data that we can retrieve. This is when we came across AutoEncoders and implemented that into a simple classifier model to work with whatever data we have.

What it does

The Model takes fewer labeled data (100) and is able to train at a relatively higher speed resulting in an overall accuracy of around 65%

How we built it

The encoder part of the autoencoder has 8 convolution layers along with MaxPooloing in between. The decoder has 4 convolution layers along with UpSampling in between. We train this model and replace the decoder part with a simple classifier model, which classifies the input images into the respective categories.

Challenges we ran into

The Autoencoder was extremely power-dependent and training required a substantial amount of GPU processing power.

Accomplishments that we're proud of

Using the whole dataset we were able to achieve around 90% accuracy and using just 100 labeled data we got around 65% accuracy.

What's next for AutoEncoder for Fashion-MNIST

We want to improve our AutoEncoder to tackle more critical tasks like Object Detection and Segmentation

Built With

• colab
• matlpotlib
• numpy
• python
• tensorflow