The Nobel Prize winner & French author André Gide said, “Man cannot discover new oceans unless he has the courage to lose sight of the shore”. This rings true with enterprises that made bold investments in cutting-edge AI that are now starting to reap rich benefits. Artificial Intelligence is shredding all perceived boundaries of a machine’s cognitive abilities. Deep Learning, at the very core of Artificial Intelligence, is pushing the envelope still further into unchartered territory. According to Gartner, “Deep Learning is here to stay and expands ML by allowing intermediate representations of the data”.

What is Deep Learning?

Deep Learning is a subset of Machine Learning that is based on Artificial Neural Networks (ANN). It is an attempt to mimic the phenomenal learning mechanisms of the human brain and train AI models to perform cognitive tasks like speech recognition, image classification, face recognition, natural language processing (NLP) and the like.

The tens of billions of neurons and their connections to each other form the brain’s neural network. Although Artificial Neural Networks have been around for quite a few decades now, they are now gaining momentum due to the declining price of storage and the exponential growth of processing power. This winning combination of low-cost storage and high computational prowess is bringing back Deep Learning from the woods.

Improved machine learning algorithms and the availability of staggering amounts of diverse unstructured data such as streaming and textual data, are boosting performance of Deep Learning systems. The performance of the ANN depends heavily on how much data it is trained with and it continuously adapts and evolves its learning with time as it is exposed to more & more datasets.

Simply put, the ANN consists of an Input layer, hidden computational layers, and the Output layer. If there is more than one hidden layer between the Input & Output layers, then it is called a Deep Network.

The Neural Network

The Neuron is central to the human Neural Network. Neurons have Dendrites, which are the receivers of information and the Axon which is the transmitter. The Axon is connected to the Dendrites of other neurons, through which signal transmission takes place. The signals that are passed are called Synapses.

While the neuron by itself cannot accomplish much, it creates magic when it forms connections with the other neurons to form an interconnected neural network. In artificial neural networks, the neuron is represented by a node or a unit. There are several interconnected layers of such units, categorized as input, output and hidden, as seen in the figure. 

The input layer receives the input values and passes them onto the first hidden layer in the ANN, similar to how our senses receive inputs from the environment around us & send signals to the brain. Let’s look at what happens in one node when it receives these input values from the different nodes of the input layer. The values are standardized/normalized-so that they are all within a certain range-and then weighted. Weights are crucial to a neural network since a value’s weight is indicative its impact on the outcome. An activation function is then applied to the weighted sum of values, to help determine if this transformed value needs to be passed on within the network. Some commonly used activation functions are the Threshold, Sigmoid and Rectifier functions.

This gives a very high-level idea of the generic structure and functioning of an ANN. The actual implementation would use one of several different architectures of neural networks that define how the layers are connected together, and what functions and algorithms are used to transform the input data. To give a couple of examples, a Convolutional network uses nonlinear activation functions and is highly efficient at processing nonlinear data like speech, image and video while a Recurrent network has information flowing around recursively, is much more complicated and difficult to train, but that much more powerful. Recurrent networks are closer in representation to the human neural network and are best suited for applications like sequence generation and predicting stock prices.

Deep Learning at work

Deep Learning has been adopted by almost all industry verticals at least at some level. To give some interesting examples, the automobile industry employs it in self-driving vehicles and driver-assistance services, the entertainment industry applies it to auto-addition of audio to silent movies and social media uses deep learning for curation of content feeds in user’s timelines. Alexa, Cortana, Google Assistant and Siri have now invaded our homes to provide virtual assistance!

Deep Learning has several applications in the field of Computer Vision, which is an umbrella term for what the computer “sees”, that is, interpreting digital visual content like images, photos or videos. This includes helping the computer learn & perform tasks like Image Classification, Object Detection, Image Reconstruction, to name a few. Image classification or image recognition when localized, can be used in Healthcare for instance, to locate cancerous regions in an x-ray and highlight them.

Deep Learning applied to Face Recognition has changed the face of research in this area. Several computational layers are used for feature extraction, with the complexity and abstraction of the learnt feature increasing with each layer, making it pretty robust for applications like public surveillance or public security in buildings. But there are still many challenges like the identification of facial features across styles, ages, poses, effects of surgery that need to be tackled before FR can be reliably used in areas like watch-list surveillance, forensic tasks which demand high levels of accuracy and low alarm rates.    

Similarly, there are several applications of deep learning for Natural Language Processing. Text Classification can be used for Spam filtering, Speech recognition can be used to transcribe a speech, or create captions for a movie, and Machine translation can be used for translation of speech and text from one language to another.

Closing Thoughts

As evident, the possibilities are endless and the road ahead for Deep Learn is exciting! But, despite the tremendous progress in Deep Learning, we are still very far from human-level AI. AI models can only perform local generalizations and adapt to new situations that are similar to past data, whereas human cognition is capable of quickly acclimatizing to radically novel circumstances. Nevertheless, this arduous R&D journey has nurtured a new-found respect for nature’s engineering miracle – the infinitely complex human brain!