In an interesting keynote speech at the SmartData Conference, Dr. Gary Marcus said statistics and Big Data are not helping the AI field, and we should go back to cognitive science to improve our approaches to AI.
Gary Marcus is:
a research psychologist whose work focuses on language, biology, and the mind. Dr. Marcus is a Professor in the Department of Psychology at New York University and Director of the NYU Infant Language Center (Wikipedia).
Dr. Gary Marcus
In his keynote titled Smart Machines – and What They Can Still Learn from People, he told us the current status of AI and what we can do to enhance it.
The following is a summary of his points.
What AI is supposed to do but cannot
Marcus started with what AI should be able to do in the first place.
- Master a wide range of tasks (note that current AI can do only one thing well, such as playing chess, operating a driverless car, or translating languages)
- Understand things like Wikipedia
- Be able to learn for itself
But he continued that the current AI cannot do any of these. Machines cannot excel humans in judgment or learning new things. He also said the current AI is a bunch of idiot savants. Paraphrasing his statement, what we need is intelligent and capable servants instead.
Marcus presented a rough idea of the size of the knowledge market in consultation with Google’s Dr. Peter Norvig. Their back-of-the-envelope calculation is as follows:
- World economy is roughly $70 trillion now
- World economy will be roughly $80 trillion in 2025
- Knowledge economy is 1/4 to 1/2 of the total economy ($20 trillion to $40 trillion)
- Smart technology will improve the knowledge economy by 2–5% ($0.4 trillion to $2 trillion)
This indicates a good opportunity for growth. With this potential, investors, enterprises, and entrepreneurs would love to jump into this market. Then, why doesn’t AI seem to be making progress as expected?
What has been wrong with approaches to AI
Marcus listed three reasons why the current AI approaches have not worked very well:
- Heavy reliance on statistics (note this guarantees correlation but not causation)
- Heavy reliance on Big Data (Marcus gave nine reasons why this does not work in one of his articles in The New York Times)
- Forgotten roots
In short, Big Data and statistics simply deal with correlation but not causation. To paraphrase, they are not getting at the root of the problem but dealing with symptoms. This is not intelligence but simple correlation. The last point is that AI is supposed to understand human intelligence and therefore cognitive science. Marcus concluded that we need to understand human intelligence to make AI better.
Critic on Numenta
In a very simple summary, Marcus concluded that Numenta‘s approach is based on Big Data analysis, and he does not buy its one-size-fits-all approach for effectiveness. It is interesting that I cannot find any rebuttal from Hawkins, even though the criticism was published in March 2013.
IBM has established a research group to work on Numenta’s learning algorithms at its Almaden research lab in San Jose, California.
But Hawkins’s company, Numenta, has made little impact on the tech industry, even as machine learning has become central to companies such as Google.
It will be interesting to observe how Numenta and IBM can improve this area.
More articles and videos
He has a few articles and videos that are quite interesting. The list is as follows.