Advances in artificial intelligence (AI) go beyond algorithms—they also include associated methods to aggregate and parse new sources of data and use it to develop new applications in an expanding range of industries. We spoke to subject matter specialists and industry experts to learn about the breakthroughs expected in each of these aspects of AI in 2021.
Speaking at the EmTech MIT 2020, in October 2020, Geoffrey Hinton, VP and Engineering Fellow at Google, expressed faith that deep learning will continue to advance as new methods, such as transformers, augment it by reading contextual information. In general, transformers take natural language processing to the next level by parsing for a vector of multi-faceted relationships.
Typically, current methods look for sequential relationships such as word embeddings or relationships suggested by syntax or grammar to capture their meaning. Transformers go further by looking at words describing context to find connections. For example, a text database mentioning a city like Paris with museums would tie modern art with the Paris Museum of Modern Art.
Natural language processing will progress by significantly reducing the ambiguity in the interpretation of language by machines. Bigger neural networks, such as Generative Pre-trained Transformer 3 (GPT-3), created by OpenAI funded by Elon Musk, are better able to emulate human intelligence, such as generating text. To be sure, even GPT-3 is only a fraction of the size of the human brain's neural network. Deep Learning will continue to progress as machines' neural networks' size comes closer to that of the human brain. (Read also: When Will AI Replace Writers?)
Transformers have similarly advanced image processing. Convolutional Neural Networks are the existing method for image processing. They do sequential processing, which would capture, for example, streets and their intersections. Transformers go to the next level and map entire neighborhoods and regions by parsing a vector of relationships.
Deep Learning had a game-changing achievement in 2020 when Google's Deep Mind constructed with its AlphaFold program the 3D structure of proteins, the building blocks of life and cells. For some time, artificial intelligence has held the promise of shortening the time for drug discovery by substituting manual laboratory experiments with analysis of data, predicting the shape and functions of proteins from their amino-acid sequences. In 2021 and beyond, more proteins will be identified and will significantly accelerate drug development and lower costs.
In the initial stages, the military used genetic algorithms to determine the flight path for drones flying thousands of miles to hit their targets.
"When you are searching for a solution in a space that is infinite, genetic algorithms (GA) find a 'reasonable' solution because we cannot possibly search every choice in the search space. GAs give rise to heuristics—the right choice when it is impossible to know the best option. We resort to GA to find a solution that is 'good enough' (a heuristic) and have an infinite or computationally infinite number of possible solutions," Gene Locklear, Senior Artificial Intelligence Research Scientist at Entrust Solutions, told us.
"GAs don't need data to operate. They arrive at solutions by a stochastic process (randomness). The configuration of the DNA of the GA is very much data-driven; the GA needs to know what are the possible components of a decision before it can randomly combine those components with seeing which one(s) are optimal."
5G networks offer numerous choices for routing traffic often generated randomly, each of them varying in the demands they make on bandwidth. "Bandwidth is the chokepoint ...never enough...GAs search for the best possible route for network traffic (best path requires the least bandwidth), which is a great place for heuristics because the number of paths is computationally infinite," Locklear said.
Service providers struggle to control the operating costs of 5G networks as data volumes explode. They will use genetic algorithms and deep learning to optimize their operations to manage traffic 2021 and beyond.
Spatial computing, the merger of the digital and our physical work and living spaces, and their representation in XReality made interactive with sensors, will have a breakout year in 2021 with the launch of Apple's iPhone 12 Pro according to David Rose, Futurist at EPAM Continuum, MIT lecturer, entrepreneur, and author of Enchanted Objects: Design, Human Desire, and the Internet of Things.
Spatialized Location and Mapping use stereo cameras or a LIDAR (Light Detection and Ranging) sensor to read geometry and superimpose content to create a blended world," Rose explained.
Apple's iPhone Pro 12 has three tiny cameras that capture depth by measuring the time it takes for structured light to bounce back. That lends a three-dimensional view with granular detail, including underwater rocks in a landscape picture of lakes.
Generative AI and its twin—discriminators—are the tools making 3D imagery. Both are variants of deep learning—one of them generates landscapes, and the other is programmed to be a judge about the fidelity of the images.
Taking an example from an engagement with a landscape design company, David Rose told us:
"Generative AI automatically selects and positions 3D trees, bushes and landscape elements over existing Google Street views. Then another neural network, called the discriminator, makes judgments on these proposed designs. The generative AI iterates again until we have a high-scoring solution."
Generative AI has gained notoriety in recent years as it has become the tool to create deep fakes, including pornographic videos of celebrities. Generative AI, termed Generative Adversarial Network, creates fake images by mining existing pictures and video databases. Discriminators tell the difference between the real and the doctored image or video. It will be challenging for discriminators to separate the real from the fake in 2021 and beyond. Transformers will make Generative Adversarial Networks so sophisticated that fake images look eerily realistic.
2021 will see increasing mining of geospatial data to improve the experience of mobile phone users. 5G Telecom heightens the quality of experience for mobile phone use. For the first time, 5G brings network slicing, enabling service providers to customize the quality of an experience in environments such as sports stadiums.
Bandwidth demands have spiked with second-screen experiences providing opportunities to see the entire stadium on a mobile device or replay excerpts of games. Service providers use artificial intelligence to find ways to monitor and adapt the quality of service to lower churn rates by meeting customer expectations.
"We bring together disparate geospatial data sources such as crowdsourced mobile devices usage and radio signal propagation data to visualize the data flow in the context of their landscape. Flat surfaces are less likely to distort radio signals than or those that are undulating or with many buildings. Correlation of data on the propagation of radio signals and service quality tells of the effect of buildings' shapes and surfaces on the degradation of signals. Service providers can prioritize maintenance or investment in assets to achieve their service quality goals."
2021 and beyond use cases will expand to include cost reduction and productivity improvement, and innovation. "AI implementations may be beginning to expand beyond efficiency….Respondents rated creating new products and services as the third-highest overall AI objective… Seasoned adopters are even more focused on this than the other maturity segments," according to Deloitte's 2020 State of AI in the Enterprise report by the Deloitte AI Institute.
Hughes Network Systems moved aggressively ahead to launch AI-enabled products and customized services for customers in 2020 and anticipates an acceleration in 2021. "Through our use of AI, we can assess a list of customer locations, often several hundred or even thousands of locations, and tell the customer the best mix of providers and circuits they should use to meet their needs and [service level agreement] SLA requirements," said Jeff Bradbury, Senior Marketing Director, North America at Hughes.
Bradbury works closely with the managed network services product and AI Center of Excellence teams.
"We've also applied AIOps to help our customers improve network performance by automatically predicting and preempting — or 'self-healing' without any human intervention — undesirable network behavior caused by WAN edge devices."
The next level involves parsing customers' data to customize services. "A store augmented to serve as an e-commerce fulfillment center will exhibit different types, volumes, and traffic patterns than other store locations. By using AI to analyze, predict, and ultimately automatically optimize the network based on their workloads and time-of-day usage, we can create a uniquely customized network based on that specific customer's usage patterns," says Jeff Bradbury. (Read also: Utilizing Visual Artificial Intelligence for Ecommerce.)
"Another new offering in 2021 will AI-power new consumption models for our customers, like removing price variation across sites based on vendor, bandwidth capacity, and transport type, and offering a flat price across sites for a common performance SLA."
Jeff Bradbury highlighted visual recognition and emotion detection in Hughes' digital signage services in the domain of emerging technologies. "While this is still a very early technology, we have customers who are piloting the use of visual cues to tailor the messages and imagery based on gender, age, even the emotions detected by the digital signage system," shared Jeff Bradbury. Emotion detecting digital signage technologies will scale as it proves itself in the high-end markets where it currently services customers.
Adoption in Enterprise
Enterprise was slow to adopt AI when we last wrote on the subject. (Read also: AI in Business: The Transfer of Expertise from Internet Companies to the Enterprise.) For 2021 and beyond, the enterprise will take a different tack to increase adoption.
"Introduction of consumer technologies like voice assistants and browsers for search greatly increases AI usability in the enterprise. Business intelligence tools for searching data from multiple databases, for example, is cumbersome. Instead, managers can search for data with voice assistants and see the results on a mobile device," explains Ajay Dawar, an investor in B2B AI ventures, and the VP of Product Management, Conga, a software company that uses natural language processing for analysis of contracts.
Readily usable AI software increases humans' participation, who can then exercise their judgment to make the best of insights gained with AI's assistance. "Leads generated by AI are not enough. Humans can draw on the company knowledgebase to pursue opportunities with the best prospects. They also intuitively decide on the outreach--the timing, the frequency, and who in the company to approach. Over time, the data predict conversion rates and the products and services most likely to elicit interest," Ajay Dawar told us. "Humans can also spot data errors when they can see it on a browser,"
As humans interact with the data, it sparks new ideas for data exploration: "Natural language searches will look at the concepts instead of individual terms. Force majeure, for example, is searched as force or majeure today. Redesigned search engines will look for any emergency, whether related to health, weather, or other black swan events," said Ajay Dawar.
"Today, search engines classify clauses in contracts by their types such as non-disclosure or indemnity. In the future, they will use this data to measure the risk exposure."
We anticipate that platforms' deployment will significantly accelerate the pace of AI adoption in 2021 and beyond. "Industry-specific platforms tailored for the needs of individual sectors are best suited for accelerated adoption," says Beena Ammanath, Executive Director of Deloitte AI Institute.
2021, by all indications, will be an inflection point where AI crosses from early evolution to the next phase of maturity with a broader range of algorithms, data sources, use cases, and adoption across industry sectors.