This repository showcases my exploration of autoencoder architectures applied to the Kuzushiji-MNIST (KMNIST) dataset. Autoencoders are a class of neural network models used for unsupervised learning, particularly in dimensionality reduction, feature learning, and data compression tasks.

An autoencoder is a neural network architecture consisting of an encoder and a decoder. It learns to encode input data into a lower-dimensional latent space representation and then reconstruct the original input data from this representation. This process is achieved by minimizing the reconstruction error, typically measured using a loss function such as mean squared error (MSE).

A basic autoencoder architecture comprising an encoder and a decoder with fully connected layers. It learns to compress and reconstruct input data without incorporating any additional complexity.

This autoencoder architecture consists of multiple hidden layers in both the encoder and decoder, allowing for more complex representations to be learned. It provides greater flexibility in capturing intricate patterns within the data.

Utilizing convolutional layers in both the encoder and decoder, this architecture is specifically designed for handling spatial data such as images. It leverages the convolutional operation to capture local patterns efficiently and is well-suited for image reconstruction tasks.

A probabilistic variant of the traditional autoencoder, VAE learns to encode input data into a probability distribution in the latent space. It introduces stochasticity during training, enabling the generation of new data samples by sampling from the learned distribution. VAE is particularly useful for generating diverse and realistic data samples.

The Kuzushiji-MNIST (KMNIST) dataset is a variant of the original MNIST dataset, containing 28x28 grayscale images of handwritten characters from the Kuzushiji script. It comprises 10 classes, representing different characters of the Kuzushiji script, making it suitable for image classification and reconstruction tasks.

I utilized transfer learning from TensorFlow Datasets (TFDS) to load the KMNIST dataset and implemented the four types of autoencoder architectures using TensorFlow and Keras. While the models may not achieve perfect reconstruction, I endeavored to apply the autoencoder concept effectively to the KMNIST dataset.