Luke Pensworth
Artificial Intelligence (AI) involves creating a computer system that can adapt its instructions according to circumstances. Such systems work on probability rather than a straightforward procedural code execution model.
Procedural language programs can come close to this concept with a cascaded IF statement. In the conditional branching model, the programmer filters out all possibilities contained in the input to execute a different piece of code according to a variable’s value, ending up at an ELSE, which leads to a default action.
AI uses probability, which forms a system of “heuristics.” Where a cascaded IF statement says “If X is Y do section A, if X is N do section B, otherwise, do section C”. An AI program would say “X is probably Y by 70 percent and probably N by 40 percent and probably Z by 20 percent otherwise it’s something else.”
The “learning” part of AI is known as “machine learning.” Deep learning is an evolution of that system. It is based on an architecture called “neural networks” and uses a technique known as “representational learning” to acquire a knowledge base.
Neural networks
Neural networks, or, more accurately, artificial neural networks, aim to model the connections that the human brain makes in order to think. A computer neural network includes a number of specialist units. Each unit stores information about one aspect of a topic – it is a module that specializes in a list of known facts. These units are called perceptrons and a typical neural network has two layers of them: one for input and one for output. Deep learning uses an evolved neural network, called a deep neural network. This structure has a number of extra hidden layers between the input and output strata.
A programmer would understand perceptrons as functions in a library. Some perceptrons have public declarations, like the utilities available in an API, whereas those hidden layers in a deep neural network have internal declarations and so serve other functions in the library rather than providing services to other programs.
A neural network links together perceptrons to produce a result. So, a shape recognition system would refer to a dimension perceptron, a scale perceptron, a rotation perceptron, and an angle perceptron to adjust the results from a reference shape database until it finds a match.
Representational learning
The “learning” function of AI is a fundamental characteristic that defines the field.
The AI system is better suited to situations where the programmer doesn’t know in advance all of the possible values of X. However, recording what happened last time adds extra weight to a possible outcome. So, if the input is Y many times, that value’s probability scoring will increase. The program records the outcome of each test. With each run, the program “learns” the likelihood of every possible outcome.
One problem with machine learning is that there needs to be a form of confirmation that a value match is correct. In some systems, this requires human intervention, which is called “supervised learning.” Deep learning uses “representational learning” to build up that database of probability scoring. Without a supervisor to confirm a successful match, the program needs a reference database.
Deep learning adds on an extra perceptron that acts as a reference. It can then search through a mass of reference data looking for matches to the input. The input data could be a text database, coordinates fed in from a probe, or an image received from a camera.
Deep learning characteristics
The key identifiers of deep learning that set it apart from machine learning lie with those hidden layers and a large reference data pool. A deep learning system goes through many more modification steps than a typical machine learning program.
Each phase in a deep learning system modifies the probability of an outcome by comparing more factors. The number of perceptrons in a neural network increases the accuracy of the system. Also, the power of the system greatly depends on the amount of reference data available. Without a very large reference source, the deep learning system has to rely on human intervention for its judgments.
Limitations of deep learning
Deep learning does not bring science closer to the typical futuristic nightmare of science fiction movies. A standard scare story is that a blank-face robot butler could suddenly realize it’s being exploited and murder the entire household. Deep learning doesn’t bring those disasters closer to reality. This is because any program of any structure is only as clever as its programmer. Computers can only emulate human nature because they cannot model spontaneity.
Most deep learning projects to date have been supervised systems, so they aren’t really deep learning implementations. Deep learning systems just match inputs to reference data. So, if you need a deep learning system to recognize a figure in a picture, its success is reliant on the reference pictures you give it. For example, if you store pictures of a fish, a dog, a bird, and a human in the system and then present it with a picture of a giraffe, it will tell you that it is probably a dog.
For each test, the humans running the AI program have to judge results accuracy and work out what extra reference data the system needs in order to produce better results. Thanks to access to large amounts of reference data, some deep learning systems are becoming very good and less reliant on supervised learning. For example, identification systems for security surveillance can perform very well if they are given the nation’s passport photo database as a reference.
Deep learning specialties
Deep learning is a good option in situations where results require a lot of testing of propositions against a large amount of data.
One area of work where deep learning is making a strong contribution is in medical research. Thanks to the genome project that lists all of the DNA in a human body and other reference information about biological behavior, deep learning systems can save a lot of time.
Another application lies with civil engineering where a lot of calculations for stresses and loads need to be performed. Modeling the performance conditions of a car, a plane, or a ship has enabled deep learning to accelerate the design phases of those high-cost products.
Deep learning is also being deployed in marketing, where it partners with big data to spot trends in consumer requirements.
Deep fakes
A big area of study at the moment in the deep learning field is the concept of “deep fakes.” This is a system that accurately depicts humans. You have probably already encountered a chatbot. This is a help support chat window that isn’t manned by a person, but by an AI program. Deep learning can make those responders seem more human and avoid getting caught in a question loop.
The biggest deep fake revelation in the news today is the manipulation of video. This enables a creator to make a real person do and say what the creator wants. This is an advancement in CGI systems used for video games and has the potential to cause real trouble by posting fake but believable videos.
