Original article was published on Deep Learning on Medium


The possibility of autoencoders has been well known in the field of neural systems for a considerable length of time, and the primary applications go back to the ’80s. Their most conventional application was dimensionality decrease or highlight learning, however more as of late the autoencoder idea has gotten all the more broadly utilized for learning generative models of information. The absolute most remarkable AIs during the 2010s included inadequate autoencoders stacked within profound neural network.

An autoencoder is a neural network that figures out how to duplicate its contribution to its yield. It has an inside which is a hidden layer that depicts a code used to speak to the info, and it is established by two primary parts: an encoder that maps the contribution to the code, and a decoder that maps the code to a remaking of the first information. It compress the input into lower-dimensional code and then reconstruct the output from this representation. The code is a compact “summary” or “compression” of the input, also called the latent-space representation. To build an autoencoder you need 3 compositions an encoding method , a decoding method, and a loss function to compare the output with the target.

The architecture

The least complex type of an autoencoder is a feedforward, non-recurrent neural network like single layer perceptrons that take part in multilayer perceptrons (MLP) — Both the encoder and decoder are fully-connected feed forward NNs. Code is a single layer with the dimensionality of your choice. The number of nodes in the code layer is hyper-parameter that we set before training the autoencoders. having an info layer, a yield layer and at least one concealed layers associating them — where the yield layer has a similar number of nodes (neurons) as the information layer, and to reproduce its sources of info (limiting the contrast between the info and the yield) rather than foreseeing the objective worth. Consequently, autoencoders are unsupervised learning models.

The working

First the information goes through the encoder, which is completely associated ANN, to deliver the code. The decoder, which has the comparable ANN structure, at that point delivers the yield just utilizing this code. The objective is to get a yield indistinguishable with the information. Note that the decoder design is commonly the perfect representation of the encoder. This isn’t a prerequisite yet is normally the situation. The main prerequisite is the dimensionality of the information and the yield should be the equivalent. Anything in the center can be played with. There are 4 hyper-boundaries that we have to set before preparing an autoencoders : code size, number of layers, number of hubs per layer, loss function.


An autoencoder is a neural network architecture capable of discovering structure within data in order to develop a compressed representation of the input. Many different variants of the general autoencoder architecture exist with the goal of ensuring that the compressed representation represents meaningful attributes of the original data input; typically the biggest challenge when working with autoencoders is getting your model to actually learn a meaningful and generalizable latent space representation.

Because autoencoders learn how to compress the data based on attributes (ie. correlations between the input feature vector) discovered from data during training, these models are typically only capable of reconstructing data similar to the class of observations of which the model observed during training.

Applications of autoencoders include:

  • Anomaly detection — Anomaly detection is the way toward discovering anomalies in a given dataset. Exceptions are the information questions that stand apart among different items in the dataset and don’t fit in with the typical conduct in a dataset.
  • Data denoising (ex. images, audio) -A denoising autoencoder is a particular sort of autoencoder, which is by and large classed as a kind of profound neural system. The denoising autoencoder gets prepared to utilize a concealed layer to reproduce a specific model dependent on its sources of info.
  • Image inpainting — Picture Inpainting is typically surrounded as an obliged picture age issue. The system must have the option to accept a setting as an info, and to yield a picture of indistinguishable measurements from the missing patch.
  • Information retrieval — Data recovery is the movement of getting data framework assets that are applicable to a data need from an assortment of those assets. Searches can be founded on full-text or other substance based ordering.