My Quest
A Puzzle: High-Performance AI without Understanding?
Some AI techniques seem to do the impossible, e.g., they translate languages or score essays with a skill comparable to humans. For example, deep learning can translate languages and LSA (latent semantic analysis) can be used for essay scoring. Yet, in both cases, there is no language understanding going on, at least not in the sense that we think of understanding. How can this be?
The simple answer is that if we treat language as data and we have enough data — or if the application does not require a fine-grained world model — then we can get by with superb pattern recognition approaches and modeling the language itself, rather the world. With enough language data such systems can acquire a good-enough language model to perform these tasks or even to generate prose in the style of certain authors.
What feels wrong about today’s approach to AI is the willful dismissal of prior work in the field and caricaturing significant contributions in knowledge representation and reasoning as “rule-based” and nothing more, which would be the same as caricaturing deep learning as “perceptron-based learning”.
Instead, each approach provides a tool, and each has limitations. Each has had its own period of wild hype, incredible media coverage, successful applications, discovery of limitations, and a realization of areas where there is brittleness.
I stay current with the latest research in deep learning, symbolic AI, and neuro-symbolic AI to provide the best possible consulting services for you. I am motivated by this desire to understand the limitations of each kind of AI — classical and deep learning — to best understand how these should be combined in future systems.