There is little doubt machine learning has become one of the most powerful technologies in the last decade. The emphasis on “learning” in machine learning allows computers to make better and better decisions, based on previous experiences.
Advances in this technology have allowed for recent breakthroughs that promote faster and more efficient business intelligence, using abilities ranging from facial recognition to natural language processing.
Machine learning programs can be thought of as individual components, or subprograms, of AI that can operate independently. The goal of true artificial intelligence (a computer or program that thinks and communicates like a human being) has not yet been achieved. However, individual machine learning programs have been trained to specialize in performing certain tasks that are quite useful. For a variety of reasons, machine learning is often referred to as AI. Combinations of a wide variety of machine learning programs, acting as subprograms, have the potential to support the goal of true artificial intelligence.
These machine learning assistants are some of the most advanced forms of artificial intelligence currently on the market. While they can’t discuss philosophy, they can understand basic commands and have a fairly large vocabulary. Virtual assistants can help with daily tasks, such as making calls, providing reminders of meetings, managing to-do lists, and taking notes. Some of the more advanced ones (Flamingo AI) can reduce research time by up to 75 percent. They augment human research by finding information within the organization’s silos.
Interactions has developed a virtual assistant that helps customers with transactions. A large number of transactions can be completed without the assistance of a human agent. Interactions.com describes their product as combining conversational AI with understanding to create impactful, positive customer experiences.
At present, chatbots are not as evolved as virtual assistants. Though their ability to understand language is comparable, their purpose is, generally speaking, to act as an informational kiosk. Their responses are more limited, providing directions to businesses within a mall, or answering the phone and offering a small selection of replies. Chatbots are used for a variety of different tasks such as phone interactions, online customer support, or assisting with online tech support. A large part of the chatbot’s growing popularity is based on the fact they are easy to deploy. As a consequence, they are often an organization’s first experience with machine learning.
ML Algorithms for Writing ML Algorithms
There are a limited number of data scientists and machine learning experts, and, unfortunately, there are simply not enough to go around. In response, programs are being developed that use machine learning to develop better machine learning. Called automated machine learning (or AutoML), it is described as “democratizing machine learning,” and allows businesses to solve complicated business problems with ML programs. The high level of automation in this process allows non-experts to use machine learning techniques and models without the expertise of a data scientist. A standard machine learning pipeline typically includes:
- Algorithm selection
- Data pre-processing
- Feature engineering
- Feature extraction
- Feature selection
- Hyper-parameter tuning
The operations listed above require considerable expertise, and it takes a fair amount of time for an experienced individual to implement them. AutoML, however, automates the process and reduces the amount of time needed. It’s faster, but not necessarily better. As an automated system, it does help to eliminate human error, and, in some cases, AutoML can develop more functional models than its human counterparts.
Natural language processing (or NLP) is concerned with programming computers to process and analyze data that has been presented in the form of human language. In other words, the computer will understand and accept verbal commands. NLP makes it much easier for humans to communicate with machines. As NLP advances, not only can we trade jokes with Siri and Alexa, but people with disabilities can access computers much more easily, and it can be used to translate for people who speak different languages. As a research tool, it can “read” the written word, sifting through huge amounts of unstructured data to find useful business information.
Graph Neural Networks
Graph neural networks are a form of machine learning that uses a neural network operating on a graph structure. Similar to graph databases, graph neural networks (GNNs) place an emphasis on relationships. Graph systems are data structures that can be used by neural nets to learn certain tasks, such as regression, classification, and clustering. Historically, GNNs have been difficult to work with. However, recent advances in network architectures, parallel computation, and optimization techniques have promoted successful machine learning within them.
Graphs use lines (relationships) to connect nodes (entities or objects). Graphs often represent real-life situations. For example, in physics the nodes represent physical elements, while edges represent the field energies expressed between them (stars and gravity/magnetic forces.) Other examples of graphs used to display real-life situations include:
- Maps: Where cities are nodes and roads are edges
- Human Relationships: Humans are the nodes and their relationships are edges
- Web Graphs: Nodes are webpages and edges are hyperlinks
- Knowledge Graphs: This shows a way to logically organize knowledge. For example, the EU can be described as “governing” Spain, in a limited sense. The EU would be represented as a node, connected by an edge, to a node titled Spain, which would, in turn, be connected to a node titled Lisbon, Spain’s capitol.
- Chemistry: Molecular graphs can be used to represent atoms (nodes) and chemical bonds (edges), to express the structural formula of different chemical compounds
In April of 2019, MIT researchers used GNNs to improve a robots’ ability to mold materials into specific shapes, and to make predictions about how solid objects and liquids interact. During training, the GNN gradually learned how the particles of different materials react and reshape. This is accomplished by calculating the different properties of each particle — mass, elasticity — and predicting the particle’s movement after some kind of deliberate change.
Variational autoencoders (or VAEs) became popular after they were used to gain cutting edge results during image recognition and reinforcement training. They have quickly become one of the most popular methods of unsupervised learning for complicated situations. VAEs are very popular because they use neural networks that can be trained to detect stochastic gradient descent (smoothness properties). VAEs have shown great promise in generating a variety of complex data models, including models for faces, handwritten digits, and physical models of scenes.
Generative Adversarial Networks
Generative adversarial networks (GANs) are still fairly new, and have the potential for some serious ethical concerns. This technology uses algorithmic architectures to build two neural networks, which are “pitted” against one another (“adversarially,” and using game theory techniques), with the goal of creating an illusionary reality.
A “generator network” maps a vector into an image or audio matrix. This output is then fed into a “discriminator network,” that learns how to distinguish between computer generated content and real content. The two networks train in tandem, and as the generator network learns better and better techniques for fooling the discriminator network, the discriminator network learns better and better techniques for recognizing artificially generated content.
The end result of using game theory dynamics is the creation of content that looks and sounds like a recording of reality, but is, in fact, fake information or fake news. While generative adversarial networks have great potential for art and political humor, they also have the potential to be used in creating fake news and advertisements. One deceitful video can do significant damage to a person’s reputation.
Image used under license from Shutterstock.com