by Alon Tvina
WASHINGTON DC - AI technology is by now the reality for most forward-thinking businesses. Early adopters across sectors are implementing and running AI systems for everything from security to marketing to customer support, and witnessing substantial returns on investment.
But as AI permeates the market, it’s facing the kind of problems that tend to occur when new technologies meet actual market demands and needs. Traction with real-life scenarios are raising pragmatic issues regarding deployment, security, transparency and more.
At the helm of AI development are companies which are already rolling out solutions designed to address these problems. While machine learning, NLP and neural networks are the current bread and butter of leading AI systems, cutting edge solutions are already incorporating the next wave of AI applications, geared towards meeting the pragmatic needs of businesses across verticals. These new technologies are supporting the demands for greater transparency, augmented security and a more agile approach to design and implementation.
Original, Holistic, Pragmatic
The concept of Authentic AI is a useful tool for gauging AI.
Authentic AI requires artificial intelligence that is Original, Holistic and Pragmatic.
Original - in the sense that the technology is developed exclusively and is not an out of the box solution, which may produce suboptimal results for the specific issue currently being tackled.
Holistic - in the sense that AI permeates the technology end-to-end, working in tandem through Dynamic AI Orchestration.
Pragmatic - in the sense that as the AI is specifically utilized for the purpose it’s serving, it will be required to digest data at its most native formats, produce and present results in the most meaningful matter, with the industry’s data modeling, prediction challenges, and goal-oriented performance in mind.
Machine performance may degrade dramatically if the original task is modified even slightly, so purpose-built AI is an advantage. Note that the concept of Authentic AI is by itself time invariant and not limited to the current state of AI achievements or progress.
Within this three-sided triangle, most recent advances in AI can be attributed to the requirement for more pragmatic systems, focused on utility and functionality. Here are some of the recent breakthroughs already being rolled out today by leading AI developers and vendors:
Explainable AI: Looking Through the Black Box
Explainable AI systems are geared at meeting the practical need of enterprises for understanding the inner logic of the “AI black box”.
The opaqueness of AI architectures, also referred to as the explainability problem or the interpretability problem, is plaguing the AI world. Often, AI networks are too complex for humans to comprehend their logic. In a corporate setting, the interpretability problem carries deeper consequences than the mere satisfaction of healthy curiosity.
Businesses need to have ways to understand and reconstruct the reasons an AI system has reached its decision. This makes good business sense - without the ability to analyze decisions they cannot be improved on - but moreover, it is frequently a legal imperative. This is especially true following the GDPR which dictates that companies deploying algorithms that substantially influence the public must create “explanations” for their models' internal logic.
When organizations rely on AI platforms for decisions that impact certain protected rights, the regulator may require them to provide the rationale for their decision. When a bank refuses to give a loan, or when a country refuses to grant entry, the individuals in question may be legally entitled to an explanation. In these cases, the fact that that’s what the AI-driven system has dictated isn’t good enough. Organizations need to be able to refer to their AI systems in order to understand the variables, and the specific weight given to them, within the decision framework.
Take, for example, a system designed to evaluate a medical condition. In order to enrich the work of a physician, any kind of evaluation must provide the specific insights which led the system to its conclusion. When this happens, the insights or evidence can be used not only to explain but also to support a specific treatment program. Today, select expert systems are already providing this ability, which in some verticals may be an imperative for the implementation of certain AI systems.
Automated Machine Learning: An Agile Approach to AI
Auto ML is a new technology supporting the practical need for a more agile approach to the design and implementation of AI systems.
The standard ML workflow is labor intensive, time consuming and costly. Initially it involves data preparation, feature engineering, and algorithm selection. Then, the model has to be developed and trained, and later tested and deployed. It’s a multistep process that often takes months to complete and requires domain knowledge, mathematical expertise, and computer science skills. By automating most of the manual modeling tasks intrinsic to the process, Auto ML creates simpler solutions that are faster and cheaper to implement.
Auto ML solutions are often better tailored to the problem at hand, and dramatically cut time to market. It’s a technology that represents a fundamental shift in the approach to ML and data science, and is accelerating the proliferation of AI into businesses across verticals. Vendors who have already adopted this technology offer their clients faster solution deployment and a more agile approach to bespoke AI systems.
Adversarial AI: Increased Security Against Malicious Intent
Adversarial AI is a cutting-edge approach to better AI security, born of the need to immunise systems against malicious attacks.
All AI solutions are geared at solving specific problem. This is a strength, but also a vulnerability. Since they operate according to a closed set of variables, their inner workings are more susceptible to handling by knowledgeable parties of interest.
With simple, unobservable manipulations, models can be taken advantage of. For example, by manipulating the input data, malicious adversaries can badly compromise the system’s output. Baking security measures into the AI system is therefore essential, especially in verticals where the system’s output carries serious consequences.
One of the strategies employed today to deal with the problems of adversarial AI is the development of ensemble learning. Ensemble methods use multiple learning algorithms working in tandem, and are therefore more robust than any of the constituent learning algorithms alone. The diversity of the ensemble learning algorithm makes it exponentially harder to game, thus reducing its vulnerability to adversity.
Know your AI
AI can deliver real value to serious adopters and can be a powerful force for disruption. According to McKinsey, early AI adopters that combine strong digital capability with proactive strategies have higher profit margins and expect the performance gap with other firms to widen in the future. One of the keys to success is the implementations of systems specifically built to solve real world problems, using the cutting edge in AI technology today.
Alon Tvina is Chief Revenue Officer at Voyager Labs. Alon is a high tech innovator and evangelist who holds a key focus on leveraging emerging technology to enable high impact growth. He comes from a rich technological background, running businesses which span digital media, big data, AI, automation, and system integration. With over a decade of experience working in cross-industry and multidisciplinary organizations, Alon’s passion for game-changing technology makes him feel right at home at Voyager Labs. He holds an MBA, B.Bus, B.CI, and professional accreditations in Project Management & Business Analysis.