The quality of an artificial intelligence system matters a lot and it is good that more regulators are coming up to offer quality assurance services. However, the question that experts need to look into with a keen eye is, can we have specialized AI testing platforms, to ensure nothing goes to chance in terms of the effectiveness of a system?
Yes, we’ve reached there, as in; we are past those days of questioning whether robots could perform a certain type of task. In fact, as of now, we have agents doing almost anything in various industries.
Amazing AI Accomplishments
Among many others, the latest top project highlights how artificial intelligence thru a platform called Systematic POacher deTecter is turning poachers into prey. Followed by an AI-based app called iNaturalist which scientists are using to discover more species. There’s also another platform christened Alice, which advices business people on how to acquire funding and other business related information, among many other models.
Ideally, there are hundreds if not thousands of AI-related systems so far, that’s why ideas, like validating the effectiveness of a model as well as approving them, is slowly becoming a new business.
Accenture Launches an Al Testing Service
Accenture, a service associated with helping engineers to come up with top-notch applications that ensure agility and speed while achieving radical productivity has announced that it has launched a new service for testing AI systems.
The service will use a technique called “Teach and Test” which is specifically designed to assist companies to build, measure, monitor, and identify reliable AI systems right from their own end, using the cloud.
Well, there is a lot to be checked on these systems but to be specific; Accenture’s “Teach and Test” aims to ensure that deployed models are giving right and accurate decisions. This comes in various levels. There is also scrutinizing of the choice of data, models, and algorithms used, as the ingredients of trained machine learning.
AI System Evaluation
During the “Test” phase, the production capacity of the system is what is focused on. Al system outputs are matched or compared to KPIs (Key Performance Indicators) and evaluated for whether the system can explain how a decision or outcome was arrived at.
Innovative methodologies and cloud-based tools are employed to monitor the models on an ongoing basis to ensure sustained performance. For example, a patent-pending stabilization method would use a special algorithm to test object recognition faster.
“Businesses are adopting AI at an astonishing rate to power up their new innovations. And with that in mind, it is critical to find reliable ways to train and sustain these systems focusing on quality, to limit unwanted effects on brand reputation, business performance, and compliance to safety,” says Bhaskar Ghosh, a group executive at Accenture Tech Services.
Literary, artificial intelligence systems need to grow, just as humans grow in education. And as with human education systems, a lot of issues relating to teaching machine learning need to be addressed. In simple words, robotic extra curriculum that can foster the understanding of right and wrong, and more so what it entails to behave responsibly (for agents like social robots) are important areas to be explored by such a service –in its “Teach” package.
“Testing AI systems comes with unique challenges. Well, it’s not enough to rely on traditional application testing, AI systems demand a limitless approach to ensure that we get accurate results.” Kishore Durg, a Growth and Strategy manager at Accenture highlighted.
In addition to that, testing doesn’t necessarily need to always come after the system is set. It can also come before the model is designed to ensure nothing goes to chance. In that case, the engineers of these systems will get an opportunity to evaluate the validity life of the data, choose algorithms as well as learning models. They will also be able to foster compliance on key areas which include ethics, balance, and safety of the system right before development.
Teach and Test’s Real Life Accomplishment
According to Accenture’s team, the technique was successfully used to teach a conversational agent recently. The client was a financial service firm’s website bot -so that it could engage users with accurate, official and unprejudiced conversations and more important, to know when it is time to refer cases to a human assistant.
In most cases, the amount of time it takes to teach an agent can be a hurdle but researchers are trying to tackle that with chips like the memristors. However, in this case, with Accenture’s technology, the involved team claims that the agent was trained 80% faster than it was initially possible, not to mention that the model recorded an 85 percent mark on customer recommendations.