Artificial Intelligence Methods
Artificial intelligence systems can be divided roughly into two main categories: symbolic learning, and machine learning.
Symbolic Learning
Sometimes called GOFAI ("Good Old-Fashioned Artificial Intelligence"), symbolic learning was the first type of artificial intelligence system developed. The era of symbolic learning in AI research ran from roughly the mid '50s to the late '80s. Underlying symbolic learning is the assumption that the world can be represented as symbols, which can then by processed according to logical processes such as if-then statements. Symbolic learning was used in highly structured logical problem-solving (such as playing simple games like tic-tac-toe) and for simple robotics in which machines perform routine tasks, like on an assembly line. AI based in symbolic learning is confined only to those situations in which variables and outputs are strictly delineated.
Machine Learning
To the extent that first generation AI relied upon symbols to simplistically represent reality and produce predictable outputs given explicit instructions, it was rather limited. For AI to reproduce the more complex and dynamic features of human intelligence, it would need to be able to process complex inputs using multiple conceptual frameworks, detect patterns, and determine the relative probabilities when there are uncertain outcomes projected. Human intelligence also has the capacity to learn from experience. AI would need to have the capacity to adjust outputs based upon patterns and inferences derived from its own performance metrics. These more subtle forms of artificial intelligence, sometimes called fuzzy logic, are produced through the use of machine learning.
One of the most basic forms of machine learning is statistical learning, which is focused on pattern recognition. Used in speech recognition and natural language processing, this AI type shows how probabilistic judgment given uncertain context is critical to demonstrating human-like intelligence. In both speech recognition and natural language processing, outputs improve as data is acquired through use, an example of AI's unique abilities to demonstrate human-like experiential learning.
Deep Learning Explained
Deep learning represents the most advanced forms of artificial intelligence programming. Deep learning systems allow for input of complex data that is processed by applying differing proportional weights to each neuronal input within convolutional and recurrent artificial neural networks (CNN and RNN, respectively). In both types of artificial neural networks, each neuron processes the information it receives, and then passes that output along to the next layer, with all neuronal inputs contributing a specific proportion to the final output. This varied weight of neurons in each layer of the artificial neural network produces a synthetic "conceptual" framework of information, which allows for output that expresses a judgment based on the weighing of all the different types of information being processed by the artificial neural network. This handling of myriad factors and information processes to produce tangible insight on complex information is foundational to the technological breakthroughs that contemporary artificial intelligence systems exhibit.
Convolutional Neural Networks (CNN)
Often applied to the advanced forms of computer vision used in today's latest technologies, convolutional neural networks were designed to mimic the information processing of the human visual cortex. While first generation computer vision systems required visual input to precisely match preset criteria for object designations (e.g., height and width relationships, shapes, etc.), with CNN-based computer vision, visual processing is based upon a wider set of parameters, and the relative weight of each parameter may be adjusted based on the specific context. This allows CNN-based computer vision applications the ability to classify objects in images from multiple angles and distances and using different focal points.
Let's take recognition of a person as an example. A head looks totally different from the front or back, and yet it is still the same object. Likewise, the different parts of a person's body look like quite different objects (compare a hand to a leg), yet they all fall under the class of the human body. Human intelligence is able to determine that an image represents a specific part of a body and compare that to the body parts of other animals to determine its true class. This demonstrates how human intelligence weighs multiple factors to form a judgment. This is exactly what convolutional neural networks allow AI applications to do in computer vision. CNNs are also used in a variety of other AI contexts, including natural language processing, games, design, business intelligence (BI) analytics, and more.
Recurrent Neural Networks (RNN)
What distinguishes recurrent neural networks from other convolutional neural networks is that they are not strictly feed-forward artificial neural networks. Processing doesn't flow exclusively from input to output through the neural network layers. Instead, feedback loops are built within the recurrent neural network structures for layered information processing that helps contextualize the output of each layer based upon the information received from the previous. This duplicates in some ways how the human mind structures thought, creating a synthetic short-term memory that allows information to persist through the information processing stages and dynamically and contextually influence the output at each stage. In speech recognition programs, recurrent neural networks allow applications to adjust to accommodate the unique features of specific voices, and in natural language processing, RNNs help ensure that words are interpreted based on the broader phrases in which they occur.
Long Short-Term Memory (LSTM)
Building on the inherent promise of recurrent neural networks, long short-term memory units enhance memory capacities when information processing must move through a larger number of layers. When recurrent neural networks are comprised of very large numbers of neural layers, the recursions back to the earliest layers in the information becomes increasingly difficult to process. By categorizing information within the artificial neural network as either short-term or long-term memory units, long short-term memory networks solve this problem. Using long short-term memory units, RNNs loop information back into the layered processing patterns selectively and only as needed. Long short-term memory has been crucial to the advancement of a variety of artificial intelligence systems, including: speech recognition, grammar learning, sign language translation, sign language translation, business process management, and more.
Reinforcement Learning
Reinforcement learning is used to refine the output of artificial neural networks based upon the relative success of their functioning by adjusting the relative weight of the various neuronal inputs to influence improved information processing. Through this process, machine learning can grow with experience. Reinforcement learning can be applied to convolutional and recurrent neural networks. Google's AlphaGo program is a great example of how powerful reinforcement learning can be in the development of advanced artificial intelligence systems. Google put the AlphaGo program through thousands of Go games to adjust the weighting system of its neural networks and produce optimal results. AlphaGo ultimately beat a grand champion in Go in four out of five games based purely on its experience of playing Go games versus itself.
About ArtificialIntelligenceB2B.com
ArtificialIntelligenceB2B.com is an AI software development company committed to making custom artificial intelligence solutions that are more affordable and accessible for a broader range of B2B businesses. Founded as a division of DDA, ArtificialIntelligenceB2B.com stands on a 25+ year history of continuous innovation in advanced custom software design and development, from eLearning platforms and virtual medical simulations to medical research and diagnostic platforms, advanced eCommerce platforms, online education and career research and guidance centers, and more. Navigate through this website to learn about the many specific B2B businesses that will benefit from AI and the kinds of AI solutions that can be developed, or if you have a specific need in mind you would like to discuss, contact ArtificialIntelligenceB2B.com today.
