Limitations of AI: Data hungry, opaque, brittle, self-reliant systems

Limitations of AI: Data hungry, opaque, brittle, self-reliant systems

Though we preach that AI investments can transform businesses, we are also not naive in our beliefs in AI’s current capabilities. Most modern AI systems suffer from common issues highlighted by respectable publications that we will collect here:

Reliance on large volumes of data

Impacts deep learning algorithms. Sadly, even when data is available, it’s likely to suffer from bias.

Research on one shot learning is an attempt to solve this problem.

Reliance on labeled data

Limits supervised learning algorithms to relatively few problems where labeled data is either available or where the solution is so valuable that companies invest in preparing semi-manually labeled data.

As a response, unsupervised learning algorithms are being improved.

Limited ability to adapt

Small changes to the problem can stop a perfectly working system and require it to be tweaked by experts to get back to functioning again.

Transfer learning is an area of active research to counter this issue. Recently AlphaZero was able to master chess, shogi and Go in a relatively short time.

Lack of transparency

Impacts deep learning algorithms. Since deep neural networks are complex mathematical structures, their logic can not be easily summarized to humans.

Local interpretable model agnostic explanations (LIME) and attention techniques are being developed to address the lack of transparency. This McKinsey article offers some good visualizations on how these techniques work.

Lack of methods for integrating prior knowledge

Though some AI methods solely rely on encoded prior knowledge, some like deep learning have no way of taking in summarized information from experts. This is a significant limitation while building systems that work with domains where current science can explain most phenomena accurately.

Almost all of these issues were highlighted by these sources:

  • McKinsey
  • Gary Marcus, a professor of cognitive psychology at NYU and briefly director of Uber’s AI lab.

Though there are limitations, there is still plenty of ways to leverage AI as we explain in our AI use cases guide.

Leave a Reply

Your email address will not be published.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