Artificial Des-Intelligence or Why machines will not take over the world. At least not now.

Part I: There is no Artificial Intelligence.

Illustration by Karen Zack, March 2016, https://twitter.com/teenybiscuit

It’s pattern recognition, stupid!

A friend of mine recently exclaimed that since her Siri speech recognition became much better, compared to speech recognition ten years ago, Artificial Intelligence (AI) now has the potential to rule the world. What if there is no Artificial Intelligence at all? What if the so called AI revolution is indeed an enhanced form of pattern recognition? I agree that todays pattern recognition shows better quality in recognizing patterns in language, image, orientation and similar fields. But is pattern recognition equal to intelligence, even to human intelligence?

Pattern recognition is about perception, and it is about statistical interference with a body of data. These are two areas that have become increasingly better over the past decade. Not only have businesses (like Amazon or Google) developed new techniques for distributed large scale computing using consumer hardware in large quantity. They have also developed decentralized, large scale solutions for data storage, labeled Big Data, that forms the base for more successful statistical interference. We see how both these quantitative changes have turned into a perceived new quality of enhanced pattern recognition (EPR).

Better algorithms to search unstructured information

Three factors play into the overall growth in automation. First of all, search engine technology has grown and become better in sifting through large amounts of structured and unstructured data, especially since Google introduced tools such as Mapreduce, Bigtable in the Mid-2000s, and since new Open Source Software for data mining in unstructured information collections such as Hadoop became available. Structured data is for example held in tables, where each column contains a certain kind of information (e.g. a date, the weather condition, a color) and each row represents a record. In Unstructured data in contrary you can never say, where a certain information may be held nor if it is there at all. In unstructured data, meaning is created by comparing it to other, structured data and through ranking algorithms. Unstructured information machine searching has become better.

Illustration by Karen Zack, March 2016, https://twitter.com/teenybiscuit

Attention direction through sentiment analysis

Secondly, sentiment analysis (the analysis of meaning of a succession of words) has become somewhat better, based on statistical learning techniques. This has driven advertisement platforms such as Google Adwords, analyzing for instance users emails or website contents to provide possibly related ads or Facebook’s ability to analyze user generated content streams and appropriate them as pools of attention.

For the past 200 years attention has been the driver for value generation in traditional mass media: Generating relevant content such as news or entertainment to redirect the readers attention towards advertisements. Nikolas Luhman notes that mass media does not necessarily produce consent, but it lives from debates, critique and dissent meaning, so that readers do engage with it mentally. It seems that this mixture of consent and dissent continues to create spaces of attention in individualized mass entertainment such as Facebook, Twitter and the like.

There is no learning in machine learning (not even recognizing)

Third, the area of pattern recognition for visual and auditive content has advanced in terms of the algorithms. This field has been dubbed »machine learning«, coined by IBM engineer Arthur Samuel in 1952.[1] The term on the one hand informs us, that it is about machines, and machines it is – calculation machines turned into symbol processing machines. Computers are tayloristic machines for the division of mental labor into its smallest calculable pieces.

What about the learning part? Learning is used here in a very specific and narrowed sense – in the sense of generating meaning in information through statistical inference with large amounts of existing categorized information. In practice this means, that in a first step specialists need to train a specific neural network by feeding it with massive amounts of information, such as pictures of dogs, that got actually labeled »dog« (by humans, who have assigned the meaning in a conscious act. Or you buy a pre-trained set and hope it works.

During its constitution the neural networks »looks« at these pictures, that means it processes all pixels of an image and calculates the pixels’ values to identify repetitions and similarities in color, brightness, position and such among different images, labeled »dog«. »Machine learning is very brittle, and it requires lots of preparation by human researchers or engineers, special-purpose coding, special-purpose sets of training data, and a custom learning structure for each new problem domain. Today’s machine learning is not at all the sponge-like learning that humans engage in.« (Rodney Brooks)[2]

Not only should we use the term »enhanced pattern recognition« instead of the wrong label of »artificial intelligence«. We also should no longer talk of »machine learning«, but of »machine feeding« when it comes to the training of neural networks.

It exists if you can calculate it

A given neural networks is restricted to information, that can be actually calculated, information that exists as a stream of distinct numbers. All other information is not existent to the neural network. Unlike humans, a computer neural network that is trained to process visual information can not – upon realizing it can not »recognize«, or better, process information – decide by itself, to retreat to other senses such as touching or hearing to generate a meaning from a given thing.

Microsoft’s engineers have learned this the hard way with their public »AI« Tay. When they put Tay online, equipped with neural network feeding algorithms that needed user input as a basis for functioning, the users understood: A group of trolls, techies and teenage geeks fed Tay with racist and anti-Semitic information and turned it into extremist rightwing propaganda blurter. They may have had decided to pick up on any other bias.

Because, it is, what it is: The algorithms are there, but the input training data is not only the base for recognizable patterns but also for bias. The trained neural network calculates the »meaning«, or lets better say, the statistical correlation of incoming information against the existing data body. Microsoft’s engineers switched off Tay after less then 24 hours. They tried to better it, to repair it, so it would »better recognize malicious intend«.[3] Are Microsoft’s engineers programmers, or are they educators, or are they psychotherapists?

When the industry today use the term machine learning they are willingly deceiving us, because as I have argued, there is no learning in machines.

Splitting it off

Another discussion looms behind these developments – The questioning of the premises that led to comparing the human brain with neural networks. Proponents of this idea can not think other of the human brain, than that it is an apparatus that computes. Only by this premise, they have reason to believe they could eventually produce a machine equipped with algorithms, that could replace the human brain. They are not completely wrong: Indeed, certain parts of human brain activity can be described through calculation processes similar to neural networks. But the story is far more complex.

It makes little sense to reduce thinking to the brain and simply ignore, that the whole body with its nervous system is involved in thinking. I can tell from my gut.
In addition there is not that one single mode of thinking as Kevin Kelly, editor of the Whole World Catalogue and founder of Wired reminds us: »We contain multiple species of cognition that do many types of thinking: deduction, induction, symbolic reasoning, emotional intelligence, spatial logic, short-term memory, and long-term memory.«[4]

And even within Kelly’s type-of-thinking list, there is an emptiness, that is typical for many engineers, mathematicians and the like. They are missing out on that part of the body that is dirty, that is emotional or dysfunctional. The are missing out on the effect on thinking conditioned by low blood pressure, by depression, or by desire. In short, they are missing out on those psychic and physiologic effects that make each body individual. What they leave is a premise – human brain is a computer – built on splitting off the incalculable.

Input, Output and what happens in-between

Even when reduced to logical brain functions, the degree of the brains’ complexity was shown recently when the question was turned upside down. Instead of applying neuroscience models to the human brain, scientists Erik Jonas and Konrad Kording evaluated standard models of neurosciences against a vintage microprocessor. It once sat in computers such as the Atari 800, the Apple I or the Commodore VC 20, of the early 1980s and it has the property, that one can determine every actual state that the processor is in during a calculation process. It was expected that the neuro-science models could explain what the computer was actually computing, but surprisingly they didn’t.

Jonas and Kording put in question, whether these models should then be used for researching the functioning of the human brain. Because, the actual state of the brain’s »computing« is largely unknown and it is by factors more complex than a MOS 6502 Integrated Circuit. Still they must have had fun, especially when they used the classic games Donkey Kong, Pitstop and Space Invaders to evaluate the computers »brain«.[5]

It may be possible, to produce complex neuro models that allow to compute information with outcomes, that look similar to what the brain does.

The outcome, however, for instance the decision if a picture depicts a dog or a muffin, even when it seems to be close to the human perception, does not say anything about the underlying computational processes. »Deep learning, however, produces a convolutional neural network that may not so easily reveal its weights and thresholds, nor how it has learned to ›chunk‹ the gridded input.«[6] What appears as Artificial Intelligence turns out to be a combination of algorithms and input data able to produce output that is statistically close to human perception of distinct chunks of reality.[7]

Illustration by Karen Zack, March 2016, https://twitter.com/teenybiscuit

There is no artificial intelligence

Portland based copy writer Karen Zack (twitter: @teenybiscuit) created a series of pictures in late 2015 until March 2016 that turned into memes. She was using Google image search and then rearranged the pictures of cats and biscuits, or chihuahuas and muffins using her phones album function.[8] Professor for neuroscience and psychology at Skidmore College Flip Phillips ran this meme through the image recognition algorithm of the online calculation machine Wolfram Alpha (aka Mathematica) and demonstrated a 50% hit rate and a 10–15% false alarm rate.[9] This is, where we are at with pattern recognition using day to day tools such as Wolfram Alpha.

Above I have shown, that the terms used in the field of Artificial Intelligence are often inappropriate and misleading. It would be worth, to analyze these further in regards to what kind of wishes, fears, traumata and desires they express. This analysis could be extended towards related terms such as »the cloud«, »big data«, »self-driving car«, »data mining«, »crowd-sourcing« and so on.

I have also discussed shortly the premises that reduce brain activity to something that is computable[10] and how this determines false claims about AI. It becomes obvious that we need to develop another language to talk about enhanced pattern recognition (EPR). We need an understanding what EPR means as a machine-algorithmic technology that does not lead to a Super-Intelligence, but has the potential to replace certain areas of human activity and labor.

 

Notes

[1] C.f. Roger Parloff http://fortune.com/ai-artificial-intelligence-deep-machine-learning/; More on the genesis of machine learning by Eren Golge at https://chatbotnewsdaily.com/since-the-initial-standpoint-of-science-technology-and-ai-scientists-following-blaise-pascal-and-804ac13d8151

[2] Rodney Brooks, https://www.technologyreview.com/s/609048/the-seven-deadly-sins-of-ai-predictions/

[3] Peter Lee, https://blogs.microsoft.com/blog/2016/03/25/learning-tays-introduction/

[4] Kevin Kelly, https://backchannel.com/the-myth-of-a-superhuman-ai-59282b686c62

[5] Eric Jonas & Konrad Paul Kording, http://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1005268

[6] Ken Regan, https://rjlipton.wordpress.com/2016/02/07/magic-to-do/

[7] C.f. Lance Fortnow, http://blog.computationalcomplexity.org/2017/04/understanding-machine-learning.html

[8] Karen Zack, https://twitter.com/teenybiscuit/status/707004279324696577/photo/1

[9] Flip Phillips, https://academics.skidmore.edu/blogs/flip/?p=712

[10] Ed Yong, https://www.theatlantic.com/science/archive/2017/02/how-brain-scientists-forgot-that-brains-have-owners/517599/

→ author: Francis Hunger, published on: 2017-Nov-15