Artificial intelligence has made progress but there is still much to do

Jeff Vossler, vice president at IBM Research, said the organization has reached several milestones in artificial intelligence in the past year and foresees three important areas of focus in 2019. Bringing AI-powered cognitive solutions to a platform that businesses can easily adopt is a strategic business imperative for the company, he said, as is increasing AI understanding and addressing issues like bias and trust.

A lot has been achieved in the last year in improving the understanding, accuracy and scalability of artificial intelligence, but in 2019 efforts will be focused on eliminating hentai and making decision-making more transparent.
Jeff Vossler, vice president at IBM Research, said the organization has reached several milestones in artificial intelligence in the past year and foresees three important areas of focus in 2019. Bringing AI-powered cognitive solutions to a platform that businesses can easily adopt is a strategic business imperative for the company, he said, as is increasing AI understanding and addressing issues such as bias and trust.

Regarding the advancement of artificial intelligence, Vossler said there has been progress in several areas, including speech understanding and image analysis. Work on IBM's Project Debater has been able to extend the AI's current speech understanding capabilities beyond simple question-answering tasks, allowing machines to better understand when people make claims, he said, and taking them beyond just "searching on steroids." One of the scenarios involved presenting a question that has no absolute answer - should the government increase the funding of telemedicine.

As critical as making AI better understand what's being said, progress has been made in making it recognize what it sees faster and more accurately, Vossler said. Instead of needing thousands or perhaps millions of tagged images to train a visual recognition model, IBM has shown that AI can now recognize new objects using just one example as a prompt, making AI scalable.

IBM Research's AI demonstrated a machine-listening capability for argumentative content stemming from their work on Project Debater, pictured with professional debater, Dan Zafirir, in San Francisco. (Image credit: IBM Research).

Another way AI learning becomes scalable is to have AI agents learn from each other, Vossler said. IBM researchers have developed a framework and algorithm to enable AI agents to exchange knowledge, thereby learning significantly faster than previous methods. Plus, he said, they can learn to coordinate when existing methods fail.
"If you have a more complex task, you don't necessarily have to train a larger system," said Wessler. "But you can take separate systems and combine them so that they perform this task."

Progress is also being made in reducing the computational resources needed for deep learning models. In 2015, IBM described how deep learning models could be trained using 16-bit accuracy, and today 8-bit accuracy is now possible without compromising model accuracy across all major AI data categories, including image, speech, and text. The scale of AI can also be achieved through a new technique for searching neural architecture which reduces the heavy lifting required to design a network.

All this progress should be tempered by the fact that AI must be reliable, and Vossler said that there will be a lot of focus on this next year. Like any technology, AI can be subject to malicious manipulation, so it needs to be able to anticipate adversary attacks.

At this moment, the artificial intelligence can be vulnerable to what is known as "adversary examples", where a hacker may imperceptibly change an image used to train the computer, and classify it into any category he wishes, thus sowing chaos. IBM's research division has made some progress in this regard for an attack-agnostic metric to assess the robustness of neural networks and systems directly on how to detect and defend against attacks.

Another puzzle is that neural networks tend to be black boxes and it is not clear how they reached the decision. It's a lack of transparency that hurts our ability to trust AI systems. In the meantime, it is also important to eliminate the bias so that artificial intelligence can be relied upon more and more for decision-making, these issues are challenging.

"So far we've mostly seen people get excited about the very ability to design AI systems that will be able to perform tasks," said Wessler, "and then later they try to understand if the systems are biased or have other problems with the decisions." Vossler concluded.

More of the topic in Hayadan:

• IBM is launching a service that will warn against biases in artificial intelligence applications and systems
• My friends and rivals - IBM introduced Project Debater: an AI system for debate competitions
• IBM will jointly invest with MIT about a quarter of a billion dollars in the establishment of the Watson Laboratory for artificial intelligence research