Deep learning algorithms work by mimicking the way humans learn and process information. Unlike machine learning, deep learning algorithms can predict outcomes with little to no supervision. With mounting volumes of data, deep learning is getting increasingly closer to replicating general human intelligence.
Deep Learning, a subfield of machine learning, involves creating artificial neural networks that simulate the behavior of a human brain. These neural networks contain multiple interconnected layers, wherein each layer optimizes the prediction or classification of the previous layer. Some of the widely used deep learning algorithms you should know are Convolutional Neural Networks, Recurrent Neural Networks and Transformers.
CONVOLUTIONAL NEURAL NETWORK (CNN)
Convolutional Neural Networks are multilayered neural networks designed to extract complex features at every layer to determine the output. CNNs are mostly used for image or object recognition applications.
For example, you have a dataset containing pictures of dogs and want deep learning algorithms to identify whether a dog is present in the input picture. CNNs process these images, with the first layer detecting simple features in the image like edges and corners. The output of every layer is fed to the next one and every consequent layer will identify more prominent features like eyes, nose and ears. The final layer uses all the features learned from the previous layers to determine whether the input picture contains a dog or not.
RECURRENT NEURAL NETWORK (RNN)
A recurrent neural network can store information about the past outputs of the neural network and use this information as an input for the next computation. This property allows RNNs to learn data sequences and make smarter decisions. RNNs are used when working with time-series data like audio recordings or text.
For example, let’s say you want to design deep neural networks that can autocomplete your sentences. In this case, the task is to predict the next word in a sentence. To do so, the algorithm needs to know all the words that came before in the sentence. Every word that the RNN receives is memorized in the sequence that it is written. After processing all the words, RNN assigns a probability value to every word in its vocabulary that could complete the sentence. When trained with many different sets of inputs, RNNs will autocomplete sentences with better results.
TRANSFORMER
Similar to RNN, a transformer is a neural network designed to process sequential data like natural language, but unlike RNN, a transformer processes the data all at once. These neural networks use the mechanism of attention heads to get the context for any word in the sequence. The ability of transformers to learn dependencies in the data makes them effective for language processing applications, videos and other types of time series data.
Suppose you want to understand customer sentiment about the newest smartphone. The review says, “I really like the user interface, but the battery does not last long.” Multiple attention heads in transformer neural networks will analyze the words in the review. Every attention head outputs a vector to help the final layer understand the context and determine the sentiment. One attention head learns the positive sentiment about the user interface while the other learns about the negative sentiment on battery power. The transformer calculates the probability of positive and negative sentiments for each feature type like battery and UI.
Deep learning processes data in a manner similar to that of the human brain. With escalating volumes of data, deep learning algorithms will keep becoming more efficient and get closer to replicating the functioning of the human mind.
