Artificial Intelligence and Understanding Time
Computer Systems That Innately Grasp Time
On the 6th of September in the New Yorker an article called How to Build Artificial Intelligence We Can Trust was published. The article was written by Gary Marcus and Ernest Davis. Dr. Marcus is cognitive psychologist and robotics entrepreneur. Dr. Davis is a computer scientist. This article was sent to me by a postdoctoral fellow at the Department of Social Anthropology. It interested me for several reasons, however the main reason was the concepts it mentioned of which I will talk of one: time.
Time to write some code
It is timely that code has to be written or programming has to be done, because problems has to be solved. It is the way to engineer the world better or do the problem solving of less, quicker or better. The authors broadly state:
“In particular, we need to stop building computer systems that merely get better and better at detecting statistical patterns in data sets — often using an approach known as deep learning — and start building computer systems that from the moment of their assembly innately grasp three basic concepts: time, space and causality.”
Why does not computer systems understand time? They start and they stop, 🛑 ✋ is that not enough? It must be said that the empirical material in this article is not very impressive, a few Google searches and a search on Google’s Talk to Books is not notable as source material. However the premise of the article or suggestion that running machine learning techniques have to take time into consideration to a larger degree is an interesting one.
Stacking and continuity: On temporal regimes in popular culture
Is the perception of time important for society, does technology change this, or even lack this? I think time, space and causality sounds so easy yet hard in practice. Another professor of mine has recently studied time at the Centre of Advanced Studies. His name is Thomas Hylland Eriksen and he is teaching a subject called Digital Anthropology. His article from 2007 Stacking and continuity: On temporal regimes in popular culture is on the curriculum this year and I am taking the module. As such diving briefly into this seems of possible benefit to both the author and the reader, so I will take a reductive look.
Eriksen talked of television series that were tailored made with the assumption that the viewer would switch channel and how this had been different. He mentions the change from the slow and linear to the fast and momentary. Further to this he mentions the relationship between the book and the World Wide Web. He mentioned then that there was not a scarcity of information, but of filtering.
“… since there is no vacant time to spread information in, it is compressed and stacked in time spans that become shorter and shorter.”
In broad brushstrokes Eriksen talks of music to topics of Bourdieu or economic theories such as diminishing returns. Faster is the general trend and slowness is a minor counterpart. Funnily enough he mentions that academic and conference papers are being put together and thrown out in a hurried way too. The tendency towards an extreme compression of time is what he talks of in the end he says: “Slacking and stacking is the enemy of logic and coherence.”
Timeless
Can this be said for machine learning techniques or the vast amounts of data we gather to gain insight from or perhaps the opposite?
“No computer has ever been designed that is ever aware of what it’s doing; but most of the time, we aren’t either.” -Marvin Minsky
We still have time, in a sense, but time for what? You cannot go through all the world of academia or learn everything although there are dreams of this from Elon Musk and others to be developing implantable brain–machine interfaces that can make this possible. It becomes a sense of gathering or aggregating information to do something significant or become more through this pursuit.
A combination of excitement and fair perhaps chasing to make a backup copy of humanity. It is challenging to discuss time or timelyness – time left to reach goals of sustainability. It becomes lost in tomorrow, today and yesterday. We grow old. In between these concerns of time we are occupying ourselves with understanding the ways in which machines understand time.
The Crystal Oscillator
Although it sounds like the name of a sci-fi novel or magical fairytale each digital device that tracks time will often have a crystal oscillator built into the hardware. It is an electronic circuit that produces waves to create an electrical signal with a precise frequency to keep track of time. That is how we made machines understand how we relate to time and convey information timely or make us aware of the present (not being present of course).
Quartz crystals are manufactured for frequencies from a few tens of kilohertz to hundreds of megahertz. More than two billion crystals are manufactured annually. As such time is manufactured and prepared for shipping. Does time have next day delivery? It is hard to understand whether it is relevant right now. It gives written in stone a double meaning.
What is the right time?
If you want to move into one specific application of time you may want to read Pourya Time Series Machine Learning Regression Framework. He has written an article that is both funny and attempts explaining the practical approach to this topic. It moves into forecasting – prediction is a typical approach in machine learning (if not most of ML).
I will not go in-depth into this, however it is often a question of saving time, minimising time for a specific time or timing (when) in machine learning. Amazon, Google and Facebook wants to know when to provide you with news — when is it the most relevant to you in a given moment. They additionally want to know when the timing is best to sell you products. As such measuring your time is important.
Even closer to your body a typical example is the Apple Watch or other watch producers. If they can get a better overview of your health data and understand who is more or less likely to die this can change prices of insurance. As such your time on this planet can be quantified too.
It is hard for the field of artificial intelligence to make machines understand time, however it will be possible to find ways to make services or processes more timely. Is the ultimate point for a machine learning algorithm to have a full overview of time? What if it did? If seamless integration is a goal unto itself then it would be great to consider a scenario whereby that was the case.
If we cannot trust machine learning techniques on the basis of its lack of understanding of time, causality and space we should probably reevaluate what we think should or can be achieved. The parameters for classifying what is necessary to understand is not always agreed upon. The authors of the article in New York Times say:
“Without the concepts of time, space and causality, much of common sense is impossible. We all know, for example, that any given animal’s life begins with its birth and ends with its death; that at every moment during its life it occupies some particular region in space; that two animals can’t ordinarily be in the same space at the same time; that two animals can be in the same space at different times; and so on.”
They talk of a set of background assumptions that has to be able to be made by an application in artificial intelligence. Right now this may seem unlikely, however it may be a question of time until this basic reasoning can be done easily by a machine. Then a general purpose robot such as described may become a reality for those dreaming of such a thing.
What indeed is the right time? For now I will say it is time to stop writing.
This is day 98 of #500daysofAI. My current focus for day 50–100 is on AI Safety. If you enjoy this please give me a response as I do want to improve my writing or discover new research, companies and projects.
