Deep Learning is an algorithm inspired by the way the human brain works. Originated as early as 1956, it has been achieving extraordinary results in the past decades, and this is only the beginning of what the future holds.
Computer Vision
Computer Vision is one the main domain of application for Deep Learning. There is a wide range of different usages for tasks such as:
* Image Classification (possibly also with localization) * Object Detection, for example from drones or satellites images * Object Segmentation
In the industry, it is for example used for self-driving cars.
Since 2011, Deep Learning algorithms achieve better results at traffic signs recognition than humans do.
Image generation
As a side domain of Computer Vision, we find applications such as: * Image Style Transfer : for example, to transform pictures into painting in the style of famous artists. * Image colorization * Image Reconstruction * Image Super-Resolution * Image Synthesis
This leads to amazing and scary things such as what is now known as “Deepfakes” for images or videos, even real-time in some experiments. Welcome to the new unreal world !
Natural Language Processing (NLP)
The model GPT-2 from OpenAI is a transformer-based model trained on 8 million web pages. It can predict the “next word” in a sentence. And it can be used for a number of NLP related activities such as translation, transcription, text-to-speech and even some more or less human-realistic text-generation in game environments. See this fun application.
Deep Learning is also used for advanced text-to-speech, for example to clone a voice and synthetise any text, in a realistic and natural way, after being trained with only 5 seconds of the original speaker recording.
Medicine
One application is diagnosis of radiology images (CT, scanner, etc.). Processing of text data (see natural language processing) could also help to diagnose diseases. For example, Deep Learning techniques are used to segment, classify and predict brain tumors images.
Biology
According to this article from Nature, Deep Learning is a powerful tool to classify biological data. It can perform tasks such as identifying a common type of genetic variation (single-nucleotide polymorphism). What is interesting is that the genomic information is first transformed into an image, then processed by Deep Learning algorithms.
DeepMind’s AlphaFold has made in the past years a scientific breakthrough with its very accurate prediction of proteins structures.
Recommendation systems
Recommendation systems are probably the application of Deep Learning we use most often in our daily lives. Facebook, Netflix, Amazon, Google: all the media giants make use of recommendation systems with Deep Learning.
Playing Games
AlphaGo and its successors (AlphaGo Zero, AlphaZero, etc.) from the company DeepMind (acquired by Google), uses a Monte Carlo tree search algorithm to find its moves based on knowledge acquired with Deep Learning. AlphaZero is considered the best Go “player” in the world. The south korean champion Lee Sedol, number 1 in the world in the late 2000s, and the only player to have beaten one game against AlphaGo, retired end of 2019 and declared that AlphaZero is an entity that cannot be defeated by a human anymore.
Also from DeepMind, the program AlphaStar has mastered the game of StarCraft II, and can beat 99% of the players in the world. AlphaStar uses a deep neural network that is trained from raw game data by supervised learning and reinforcement learning.
Robotics
Robotics is also a great domain for Deep Learning. One example is Imitation Learning, for example for grasping objects (researchers have been doing programming by demonstration for the robots). See this example on Teaching a Robot to Grasp Real Fish by Imitation Learning. And as for any task that is based on learning, Deep Learning algorithm are ideal for this task.
Other applications
In other applications, there are also Forecasting - Deep Learning is applied on Time Series: financial, logistical, etc. For example, forecasting of solar irradiance and photovoltaic power.
To deep dive into Deep Learning, I encourage you to follow the FastAI course and read Deep Learning for Coders with fastai & PyTorch from Jeremy Howard and Sylvain Gugger.
Jeremy gave also a fantastic TED talk in 2014 to present the wonders of Deep Learning. Since then, innovation has continued to be spectacular.