Overview of Deep Learning

Deep Learning provides human-like multi-layered processing as compared to shallow architecture. The fundamental idea of deep learning is to apply hierarchical processing using multiple layers of architecture.

And all the layers are arranged hierarchically. After numerous pre-trainings, each layer’s input data goes to its adjacent layer. Oftentimes, the pre-training of a selected layer is performed in an unsupervised way.

Deep learning aids the processing and arrangement of the data into separate layers according to its scale, its time (occurrence), or nature.

Deep learning is usually linked with artificial neural networks. Given are three categories of deep learning architectures:

• Generative architectures
• Discriminative architectures
• Hybrid deep learning architectures

1) Generative architectures- Generative architectures focus on the pre-training of a layer in an unsupervised way. It eliminates the difficulty of training lower level architectures, relying on the previous layers.

You can pre-train each layer and later include it in the model for further general tuning and learning. Doing this saves us the time of training neural network architecture with multiple layers and enables deep learning.

2) Discriminative architectures- Neural network architecture can have discriminative processing ability by having various information combinations and thus forming deep learning architecture or by heaping the output of each layer with the original data.

Also Read: 18 Time Series Analysis Tactics That Will Help You Win in 2020

3) Hybrid deep learning architectures- The hybrid architecture is a mix of the properties of the discriminative and generative architecture.

It wouldn’t be wrong if we call Deep learning a “Technique for Implementing Machine Learning”. It’s a subset of machine learning. Deep learning uses neural networks and imitates the network of neurons in a brain. This allows the machines to make accurate decisions without the involvement of humans.

However, deep learning is regarded as the evolution of machine learning but in deep learning, the computer/machine uses various layers to learn from the data, and the depth of any deep learning model is denoted by the number of layers it has. We have clearly understood what is deep learning, now it’s time to have a look at the types of deep learning.

Types of Deep Learning Algorithms

1) Convolutional Neural Networks (CNNs)- CNNs are also known as ConvNets, contain multiple layers, and are mainly used for object detection and image processing.

CNN’s are extensively used to process medical images, identify satellite images, forecast time series, and detect anomalies.

2) Long Short Term Memory Networks (LSTMs)- LSTMs can learn and memorize long-term dependencies and recall past information for long periods.

LSTMs can retain information over time. They are mostly used in time-series prediction because they remember previous inputs. LSTMs have a chain-like formation where four interacting layers communicate uniquely. LSTMs are typically used for music composition, speech recognition, and pharmaceutical development.

3) Recurrent Neural Networks (RNNs)- RNNs have connections forming directed cycles, which allow the outputs from the Long Short-Term Memory Networks (LSTMs) to fed as inputs to the current phase.

RNNs are commonly used for time-series analysis, image captioning, handwriting recognition, natural language processing, and machine translation

4) Generative Adversarial Networks (GANs)- GANs are generative deep learning algorithms that create new data cases that match the training data.

GANs have two components: a generator, to generate fake data, and a discriminator, that can learn from that false information.

They are usually used to simulate gravitational lensing for dark-matter research and improve astronomical images.

5) Self Organizing Maps (SOMs)- SOMs,  invented by Professor Teuvo Kohonen enables data visualization to diminish the dimensions of data through self-organizing neural networks.

Data visualization tries to solve the problems that humans cannot easily visualize in high-dimensional data. Self Organizing Maps (SOMs were created to help users understand the high-dimensional information.

6) Deep Belief Networks (DBNs)- Deep Belief Networks (DBNs) consists of multiple layers of latent and stochastic variables. The latent variables have binary values. They are also called hidden units.

DBNs are a heap of Boltzmann Machines with connections between the RBM layer communicates and layers, with both the previous and subsequent layers. DBNs are used for video-recognition, image-recognition, and motion-capture data.

7) Restricted Boltzmann Machines( RBMs)- RBMs are stochastic neural networks that learn from a probability distribution across a set of inputs.

This deep learning algorithm is used for classification, dimensionality reduction, collaborative filtering, regression, topic modeling, and feature learning.

Restricted Boltzmann Machines consist of two layers:

• Visible units
• Hidden units

All visible units are inter-connected to each hidden unit.

Conclusion:

Deep learning has evolved a lot in the past few years, and deep learning algorithms have become extensively popular in various industries.

Hope you got to learn a few things about deep learning. Deep learning is going to change the way we look at artificial intelligence and it is going to speed up the growth of Artificial intelligence for sure.

The post Understanding Deep Learning & its types first appeared on Yogesh Gaur.