Idea 8: Expand the Model
The final method for improving the robustness of AI
systems to the unknown unknowns is to expand the
model. We can all point to examples of ways in which
our AI systems fail because they know so little. While
it is impossible to create a “model of everything,” our
existing systems fail primarily because they have a
model of almost nothing.
There have been some notable efforts. Doug
Lenat’s decades-long effort to create a large commonsense knowledge base, CYC, led to some interesting
applications and insights (Lenat et al. 1990) and has
been licensed to Lucid (Knight 2016). Recent work
has seen the development of systems that can extract
concepts and properties from the World Wide Web to
grow and populate a knowledge base (Mitchell et al.
2015). NIST has been operating a knowledge base
population competition to evaluate such systems
(Surdeanu and Ji 2014). Google employs a knowledge
graph that contains millions of objects and relationships, and other companies including Microsoft, Yan-dex, LinkedIn, and Baidu have built similar semantic
There are some risks to expanding our models.
Every time we add something to the model, we may
introduce an error. Inference can then propagate that
error. The result may be that the expanded model is
less accurate and less useful than the original model.
It is important to test our models continually to prevent this from happening. A beautiful aspect of the
application of knowledge bases in web search is that
because millions of queries are processed every day,
errors in the models can be identified and removed.
AI has been making exciting progress. The last two
decades have seen huge improvements in perception
(for example, computer vision, speech recognition),
reasoning (for example, SAT solving, Monte Carlo
Tree Search), and integrated systems (for example,
IBM’s Watson, Google’s AlphaGo, and personal digi-
tal assistants). These advances are encouraging us to
apply AI to difficult, high-stakes applications includ-
ing self-driving cars, robotic surgery, finance, real-
time control of the power grid, and autonomous
Figure 15. Example Output from the Berkeley Image-Captioning System.
“a black and white cat is sitting on a chair.”