Deep learning is a concept designed to enable a computer system to learn by example. Developing its learning on its own, the machine acquires great reasoning skills that computers did not have before. This process is a great advance in the field of artificial intelligence. Here's what you need to know about deep learning.
Translate deep learning
Deep learning is a concept belonging to machine learning and whose objective is to enable a computerized machine to learn by example. This new mechanism allows us to understand artificial intelligence in a completely different way.
Designed based on a network of artificial neurons, the machine has the ability to analyze data received, decode it as information and interpret it in comparison with previous stored data. Its scope of application is enormous and allows artificial intelligence to develop its potential for the future.
How Deep Learning Works
The complexity of a system using deep learning varies according to its number of layers of artificial neurons, which can increase from ten to a hundred. The more these layers, the more complex the operation will be.
This allows a computer system to recognize letters, words, or even text entirely. It is through this process of deep learning that machines can recognize a face in a picture, or even recognize and differentiate animals from each other.
To identify a dog in an image, the system treats the information received as follows:
- Each layer of the artificial neural network processes a certain aspect of the image;
- At each stage the system deals with a possible answer, if it is "wrong", the program returns to the lower level until it finds the "right" answer. Once it is found, another layer of the neural network takes over and processes potential responses, etc.;
- Once the program reorganized all the information and identified the image as showing a dog, the system will have learned this concept and will be able to recognize dogs autonomously.
With deep learning, no developer needs to code programs for the computer to recognize a dog from another animal species. It is from the raw data sent (many pictures of animals in this example) that the machine will identify the canids by itself.
The quantity of raw data previously infused into the machine is very important. The larger the computer database, the faster and more efficient its learning will be.
Differences between Deep learning and Machine learning
Machine learning is a concept whose principle is to learn from a machine automatically. The computer will therefore be able to learn from the previously processed data and will then be able to give an appropriate response to a complex situation.
Deep learning is a sub-category of machine learning... However, when machine learning requires prior data processing to allow the computer to classify new data, deep learning provides a stand-alone ranking of raw data. Thus, deep learning brings many more possibilities and allows learning more similar to a human.
Application area of deep learning
You may not know it, but in your daily Internet activities you are already using deep learning technology.
When using Apple's facial recognition or Google's voice recognition for example. Here are some other examples of areas where deep learning is used:
- Medical diagnosis;
- Personalized recommendations;
- Automatic moderation of social networks;
- Financial predictions;
- Space exploration;
- Fraud and malware detection;
- The development of smart cities;
- The use of IOTs (Internet-connected objects)
Current Limits of Deep Learning
Some machine trials using deep learning have revealed to be very problematic, such as Google's artificial intelligence which in 2015 had associated an African-American user with a gorilla.
The problem was quickly corrected, indicating that the error was the reason for a lack of diversity in the photos given to the machine for training. The error therefore came from a human bias.
In the same idea, in 2016, Microsoft launched a chatbox using deep learning that could learn through Twitter exchanges with internet users.
As a result, in less than 24 hours, the robot made hateful, racist and also homophobic remarks. This is also due to relationships with engaged users.
So the problem is human bias.
IBM has even identified more than 180 biases that can have an impact on machine learning.
Lack of common sense
The deep leaning is sorely lacking in common sense. Despite its network of elaborate neurons, the system is unable to anticipate scenarios familiar to a human being. One of the best examples is that of the bus, a human being will know by observing a bus at the stop that pedestrians could cross the roadway.
It is not possible today to equip artificial intelligence with these potential scenarios using deep learning. These scenario elaborations come from a life experience and a common understanding of the world.
Lack of ethics
The development of deep learning does not consider emotions and how they affect the decisions made by a human being. The test of the first autonomous car in the United States is a concrete example.
During this test, a human driverless car moved long distances and struck a pedestrian. After analyzing the data collected from the car, the engineers realized that the smart car had faced a dilemma: stop to let a pedestrian pass and risk a pile-up, or not stop even if it took the life of a human being.
By choosing the first option, the car made a choice that was sorely lacking in ethics. This example shows the big problem of this technology today, machines are not equipped with emotions, but these are the emotions that the majority of humans uses to make choices.
It is currently impossible to let a car take to the road without a driver, or to let a computer make choices that impact the lives of individuals.
Lack of nuance
Deep learning creates an elaborate neural network, but still very succinct in comparison with the functioning of the human brain. One of the reasons is that we do not yet know all the mechanisms for the development of a cognition that takes into account rationality, emotions, but also the understanding of nuances.
Not everything is black or white and humans can discern shades of grey unlike a machine. The most convincing example is with moderation systems like on YouTube. The program removes unwanted videos such as having a hateful purpose, plagiarizing, etc.
Frequently, the system "gets it wrong" by deleting videos because it uses words associated with undesirable topics. Nevertheless, these words have taken them out of context and do not express the subject or message of the videos. This lack of nuance and understanding demonstrates the flaw in deep learning. It is therefore not currently possible to allow machines to moderate without human control afterwards.
Deep learning presents a huge advance in the field of artificial intelligence. The development of complex neural networks allows computers to learn by example and reason for themselves, depending on their learning.
For you too to make the most of your data, don't hesitate to call on Ryax!
The Ryax Team.