Deep learning, an advanced artificial intelligence technique, has become increasingly popular in recent years thanks to abundant data and increased computing power. It is the main technology behind many of the applications we use on a daily basis, including the translation of online languages and the automatic social media face tag.
This technology has also been useful in healthcare: earlier this year, computer scientists at the Massachusetts Institute of Technology (MIT) used Deep Learning to create a new computer program for detecting breast cancer.
Classic models required that engineers manually define the rules and logic for detecting cancer. However, for this new model, scientists submitted a Deep Learning Algorithm 90,000 full-resolution mammogram scans of 60,000 patients, revealing common patterns between scans of patients with and without breast cancer. It is able to predict breast cancer up to five years in advance, which is a significant improvement over previous risk prediction models.
What exactly is machine learning?
Deep learning is a subset of machine learning, a branch of artificial intelligence that configures computers to perform tasks through experience. In contrast to classical, rule-based AI systems, machine learning algorithms develop their behavior by processing annotated examples, a process called "training".
For example, to create a fraud detection program, train a machine learning algorithm with a list of banking transactions and their potential consequences (legitimate or fraudulent). The machine learning model examines the examples and develops a statistical representation of common features between legitimate and fraudulent transactions. If you subsequently provide the data for a new bank transfer to the algorithm, it will be classified as legitimate or fraudulent based on the patterns it has extracted from the training examples.
The rule of thumb is that the data is of higher quality The more accurate a machine learning algorithm is in executing its tasks, the more accurate it is.
Machine learning is particularly useful in solving problems where the rules are not well defined and can not be coded into individual commands. Different types of algorithms are characterized by different tasks.
Deep Learning and Neural Networks
While classical machine learning algorithms solved many of the problems rule-based programs faced, they can not handle soft data such as pictures and videos, audio files, and unstructured text.
For example, according to AI researchers and data, the creation of a breast cancer prediction model using classical methods of machine learning would require the efforts of dozens of professionals, computer programmers, and mathematicians, scientists Jeremy Howard. The researchers would have to do a lot of feature engineering, a tedious process that programs the computer to find known patterns in X-ray and MRI scans. Afterwards, the engineers use machine learning in addition to the extracted features. Creating such an AI model takes years.
Deep learning algorithms solve the same problem using deep neural networks, a type of software architecture that is inspired by the human brain (although neuronal networks differ) biological neurons) , Neural networks are layers of variables that adapt to the characteristics of the data they have been trained on, and are capable of performing tasks such as classifying images and converting speech to text.
Artificial Neural Network (source: Wikipedia)
Neural networks are particularly good at finding common patterns in unstructured data independently of one another , For example, when you train a deep neural network on images of different objects, you can find ways to extract features from those images. Each layer of the neural network recognizes specific features such as edges, corners, faces, eyeballs, etc.
Top Layers of Neural Networks recognize general features. Deeper layers detect actual objects (Source: arxiv.org)
The use of neural networks makes deep-learning algorithms superfluous to feature engineering. In the case of MIT's Breast Cancer Predictive Model, the project, through in-depth learning, required less effort from computer scientists and professionals, and development took less time. The model also found features and patterns in mammogram scans that were overlooked by human analysts.
Since the 1950s, neural networks exist (at least conceptually). Until recently, however, they were largely dismissed by the AI community because they needed huge amounts of data and processing power. In recent years, the availability and affordability of storage, data and computing resources has brought neural networks to the forefront of AI innovation.
What is Deep Learning used for?
There are several areas where deep learning is used to help computers solve previously unsolved problems.
Computer Vision: Computer vision is the science to use software to understand the content of images and videos. This is one of the areas where deep learning has made great progress. In addition to breast cancer, deep learning image processing algorithms can detect other cancers and help diagnose other diseases.
However, Deep Learning is also anchored in many of the daily applications. Apple's Face ID uses deep learning, as does Google Photos Deep Learning, for various features such as finding objects and scenes, and correcting images. Facebook uses Deep Learning to automatically tag people on the photos you upload.
Deep learning also helps social media companies automatically identify and block questionable content such as violence and nakedness. And finally, deep learning plays a very important role in enabling self-driving cars to understand their environment.
Voice and Speech Recognition: When you send a command to your Amazon Echo Smart Speaker or your learning-intensive Google Assistant algorithms convert your voice to text commands. Some online applications use Deep Learning to transcribe audio and video files. Google has recently released a real-time Gbox voice transcription smartphone app on the device, which you can use to learn profoundly while speaking.
Natural Language Processing (NLP) and Generation (NLG): Natural Language Processing The science of extracting the meaning of unstructured text was a historical pain point for classical software. It is virtually impossible to define all the different nuances and hidden meanings of the written language with computer rules. However, neural networks trained with large amounts of text can perform many NLP tasks accurately.
Google's translation service saw a sudden increase in performance as the company switched to deep learning. Smart Speakers use deep-learning NLP to understand the nuances of commands, such as: For example, the different ways you can ask for weather conditions or directions.
Deep learning is also very efficient in generating meaningful text, also known as natural language generation. Gmail Smart Reply and Smart Compose use Deep Learning to get relevant replies to your emails and suggestions for completing your phrases. A text generation model developed by OpenAI earlier this year produced long excerpts of coherent text.
The Limits of Deep Learning
Despite all its advantages, deep learning also has some shortcomings.
Data Dependency: In general, deep learning algorithms require large amounts of training data to perform their tasks accurately. Unfortunately, for many problems, there is not enough qualitative training data to build in-depth learning models.
Explanation: Neural networks develop their behavior in an extremely complicated way – even their creators struggle to understand their actions. Due to the lack of interpretability, it is extremely difficult to fix bugs and fix bugs in deep learning algorithms.
Algorithmic Distortion: Deep learning algorithms are as good as the data they were trained on. The problem is that training data often contain hidden or obvious distortions and the algorithms inherit these distortions. For example, a facial recognition algorithm that is trained primarily on images of white persons will be less accurate on non-white people.
Lack of generalization: Deep learning algorithms perform well focused tasks, but are poor at generalizing their knowledge. Unlike humans, a deep-learning model trained to play StarCraft can not play a similar game, such as playing poker. Eg WarCraft. In addition, Deep Learning can not handle data that differs from the practice examples, also referred to as "peripheral cases". In situations like self-driving cars, where mistakes can have fatal consequences, this can be dangerous.
The Future of Depth Learning
Earlier this year, the pioneers of depth learning were awarded the Turing Award, the equivalent of computer science, the Nobel Prize. However, work on deep learning and neural networks is far from complete. Various efforts to improve deep learning are in the works.
Some interesting papers include deep-learning models that are explicable or interpretable, neural networks that can develop their behavior with less training data, and Edge AI models, deep-learning algorithms, that can perform their tasks without to rely on large cloud computing resources.
And while deep learning is currently the most advanced artificial intelligence technology, it's not the ultimate goal of the AI industry. The development of deep learning and neural networks could give us completely new architectures.