The Waves of AI

Kyndi is bringing a new type of AI to market.

To explore how it’s new, let’s start with some AI characteristics using a scheme from DARPA. In this scheme, we can think about AI goals as having four categories: perceiving, learning, abstracting, and reasoning. Using these categories, we can see how AI has evolved, and where it might go next.

  • The first wave of AI (handcrafted knowledge): This wave focused very heavily on handcrafted systems able to reason about fixed domains, from toy domains such as a “block world” to rule-based systems designed to replicate expert knowledge in real-world scenarios. These systems were strong on reasoning with a little perception (replicating human senses such as vision and reading). They had no learning or abstraction of knowledge. These techniques remain extremely powerful and are often still used.
  • The second wave (statistical learning): This wave culminated in today’s deep learning and machine learning, which uses huge training datasets and huge compute power to achieve impressive pattern-matching outcomes. Andrew Ngis a global leader in deep learning and characterizes these technologies as extremely good at things that take people less than a second. The characteristics here are high on perceiving and learning, but at the expense of reasoning.

The next generation (natural language understanding)

Natural language processing has been a recurring theme throughout AI, first using symbolic techniques from the first wave, then applying statistical methods from the second wave. However, these approaches fall well short of human reading capabilities, as they miss out on easily observed elements of how we read and understand text.

A new approach is needed to add together all the elements that we use for reading. These include all four of the DARPA capabilities used in combination. When you and I read text, we use a mix of perception, of learning, pattern matching, analogy, simile and metaphor. We understand abstractions, we reason about what we read.

Steps toward developing true natural language understanding need to address at least these capabilities.

My next post explores more details on why language is difficult.

Michael
CTO, Kyndi
@michaelharries

Comments

Your email address will not be published. Required fields are marked *

Like to Know More?