Innovation
Paid Content : This paid content was written and produced by RV Studios of Red Ventures' marketing unit in collaboration with the sponsor and is not part of ZDNET's Editorial Content.

From Language to Knowledge: Intel's path to cognitive AI

Multimodal machine learning helps AI understand context and apply common sense to inputs.

The concept of artificial intelligence has been with us for a very long time. Five years before the first point-contact transistor was invented back in 1942, Isaac Asimov postulated artificially intelligent robots. It took another 40 years or so for rudimentary AI systems and inference engines to see practical use. Now, 40 years beyond that, AI has reached a level of practicality that is being applied across industries.

Today's AI has a workable mastery of natural language processing, image recognition, optimal path analysis, and other workhorse processes that can be applied to statistical and heuristic-based intelligence. But while AI is quite capable of filtering and identifying information, it still has difficulty understanding. AI technologies have become masters of data manipulation, but not of true cognition. 

Part of the problem is that current models and modalities can't scale to handle the sorts of problems that require true understanding. Parametric systems (that predict outcomes based on large numbers of inputs) like GPT-3, Switch-C, and Wu Dao have been fed billions and trillions of data points, respectively. But if the information isn't in the dataset, the knowledge can't exist. Something as simple as performing a search engine query, sifting through the results, and deriving meaning are beyond most current models.

These models also lack flexibility, adaptability, and the capacity for logical inference that's necessary to "think through" solutions for which pre-existing models or parameter fields haven't been prepared. The ability to solve new problems dynamically may seem possible, but only until the bounds of the source data are reached.

Physical scalability is also problematic. After all, while you can certainly store and access terabytes in the cloud, there is a propagation delay between an on-the-spot device that needs to make a judgement call and the enormous knowledge base in the sky.

But we're starting to enter what experts call the "third wave of AI." The first is characterized as the "describe" wave, where basic datasets were created that represented some basic knowledge in a tightly specified domain. First-wave algorithms were able to work through those datasets to provide information, but they were not able to learn. 

In the second wave of AI, which is what we're experiencing now, the watchword is predict. Systems use complex statistical models that are 'trained' on big data. Think about how home voice assistants are being trained on billions of voice requests by users across the world. That's big data meets statistics. Some learning is possible, prediction is sometimes surprisingly accurate, but there's still no reasoning.

What we're looking at in this article is the beginning of the third wave of AI, characterized as explain, where cognition, reasoning, and understanding are possible. Here, scientists are exploring multimodal learning, training models the same way humans acquire knowledge. Key to this may be what Intel scientists call the three levels of knowledge (3LK or Thrill-K).

According to Gadi Singer, vice president and director of Emergent AI at Intel Labs, "The third wave of AI is emerging as neuro-symbolic systems that combine neural networks with deep knowledge structures and symbolic reasoning. Intel Labs' Cognitive AI research is focused on building an architecture that can compactly accrue relevant knowledge and apply common sense explainable reasoning. We believe the Thrill-K architecture brings together the skills and knowledge required for higher machine intelligence."  

image2.png

Inside Thrill-K

The first level of knowledge is instantaneous knowledge. This is information that an AI system has at its immediate disposal. In practice, this might be data built right into the device or the parametric memory of a neural network.

The second level of knowledge is standby knowledge. This is knowledge built into knowledge bases and parametric sets that can be called upon by the AI to improve understanding and provide deeper knowledge.

The third level of knowledge is external knowledge. Think of this as all the knowledge that's out there on the web, in libraries, and in systems not built into the AI.

Cognitive systems using the three levels of knowledge can tap into that information, no longer limited by pre-trained parametric memory, regardless of how huge. As they examine a problem, they can pull standby knowledge into the analysis. If standby knowledge isn't enough, the Thrill-K AI builds a more comprehensive understanding using external knowledge as well.

When considering the scale and cost of these Tera-scale and Zetta-scale repositories in today's environment, however, the promise of Thrill-K seems far from reality. Intel is working on lowering these barriers, as Singer explains:

To advance to the next generation of machine intelligence, moving beyond statistical correlation is required. By creating systems that integrate multiple modalities containing structured deep knowledge, AI can grasp meaning and relationships between entities and improve common sense understanding. Refactoring information between the neural network and adjacent knowledge structures will allow substantially more capable AI while reducing cost and power.

AI that understands 

image2.png

Thrill-K based cognition opens up a wide array of use cases. Let's look at a few as inspiration:

  • Multimodal semantic search: We're all familiar with search – and its limits. But what if we could search across a wide range of modalities (speech, video, images, textures, data) with an understanding of the searcher's intent? As cognitive AI is added into search, we'll get better results.

  • Video and image search using language: Likewise, what if we can do better searching our videos and images? Many of us now have giant libraries of personal and professional pictures and video. What if we could just ask our libraries to find all media related to a project we're working on, and the search process could do that from the context in the media? 

  • Meaningful insight and communication: When an AI system is able to make associations on its own and explain its reasoning, we may see computers that can 'watch' a movie or presentation and then carry on a conversation about what they've 'seen.' 

  • AI assistants: Our AI assistants sometimes seem too smart for us, but also shockingly incapable, at times. What we really need are AI assistants that are smart enough to infer our intent, meet our needs, and respect our privacy.

  • Autonomous robots and drones: The key to drones and robots becoming truly helpful is their ability to navigate their environments and carry out duties without mishap. Getting from here to there is challenging, but better cognitive capabilities via Thrill-K can make robots and drones safer and more successful overall.

Intel researchers believe this three-level knowledge hierarchy and Thrill-K architecture are the blueprints for building practical and accountable AI solutions in the future.

"I expect third-wave AI technologies like Intel Labs' Cognitive AI Thrill-K architecture to emerge at a similar pace to deep learning, following a 10-year trajectory," Singer stated. "We foresee Cognitive AI implementations going from being nascent in 2021 to reaching material commercial use by 2025, and becoming widespread by the end of this decade."

To learn more about Cognitive AI research at Intel, please visit www.intel.com/labs.

Editorial standards