The Danger of Humanizing Algorithms

Misleading terminology can be dangerous. Machines are actually not learning

To many, 2016 marked the year when artificial intelligence (AI) came of age. AlphaGo triumphed against the world’s best human Go players, demonstrating the almost inexhaustible potential of artificial intelligence. Programs playing board games with superhuman skills like AlphaGo or AlphaZero have created unparalleled hype surrounding AI, and this has only been fueled by big data availability.

In this context, it is not surprising that the public, business, and scientific interest in machine learning are unchecked. These programs can go further than beating a human player, going so far as to invent new and ingenious gameplay. They learn from data, identify patterns, and make decisions based on these patterns. Depending on the application, decision-making occurs without or with only minimal human intervention. Since data production is a continuous process, machine learning solutions adapt autonomously, learning from new information and previous operations. In 2016, AlphaGo used a total of 300,000 games as training data to achieve its excellent results.

Every guide out there about how to implement machine learning applications will tell you that you need a clear vision of the problem it has to solve.

In many cases, the machine learning applications are faster, more accurate, and time-saving, therefore — among other benefits — shortening time-to-market. However, it will only address this specific problem with the data given.

But is this learning in correspondence to the way humans learn? No, it is not. Not even remotely.

Mystery Unsolved: What Is Intelligence?

As humans, we have an idea of what we consider smart or intelligent. However, scientifically speaking, it proves almost impossible to grasp and understand it.

There are several reasons for this, one of which is cultural. For example, in the West, being smart is associated with being quick. The person who answers a question the fastest is seen as the most intelligent. But in other cultures, being smart is related to considering an idea thoroughly before answering. A well-thought-out, contemplative answer is the best answer. Another reason is we can’t measure all aspects of intelligence.

For developmental psychologist Howard Gardner, there are not one but nine domains of intelligence. Only three of those are measured by an IQ test:

• Logical-mathematical
• Linguistic
• Spatial

However, the following six are entirely omitted by IQ tests:

• Musical
• Bodily-kinesthetic
• Naturalistic
• Interpersonal
• Intrapersonal
• Existential

Therefore, a high IQ does not mean success in life or indicate that a person has common sense or excellent interpersonal skills. Other theories on intelligence include Sternberg’s triarchic theory of intelligence.

What all psychological theories have in common is that intelligence is perceived as a broad term to describe human intellectual capabilities manifested in sophisticated cognitive accomplishments and high levels of motivation and self-awareness.

“Intelligence is a very general mental capability that, among other things, involves the ability to reason, plan, solve problems, think abstractly, comprehend complex ideas, learn quickly and learn from experience. It is not merely book learning, a narrow academic skill, or test-taking smarts. Rather, it reflects a broader and deeper capability for comprehending our surroundings — ‘catching on,’ ‘making sense’ of things, or ‘figuring out’ what to do.” — Mainstream Science on Intelligence

However, in the context of machine learning, learning is statistical and should answer the following fundamental question (per The Nature of Statistical Learning): “What must one know a priori about an unknown functional dependency to estimate it on the basis of observations?”

While Léon Bottou argues that this statistical nature is well understood and statistical machine learning methods are now commonplace, the quality of reasoning proves more elusive.

Anthropomorphism in AI

Anthropomorphism is the tendency to ascribe human characteristics to non-human objects (e.g. Bambi), and it is evident throughout the field. Just ask yourself, “What does artificial intelligence itself imply?”

If we look more closely into the statistical concepts trying to achieve reasoning, flawed terminology strikes especially hard in deep learning applications. In this subset of AI, artificial neural networks are built for algorithms to learn from vast amounts of data. They are centered around building a learning machine to accomplish a valuable task. It was created to summarize complex information into tangible results, inspired by the human brain. The advantage of these networks is the profound abstraction of relations between input data and the abstracted neuron values with the output data done through several layers of the networks (while traditional neural networks only contain 2–3 hidden layers, deep networks can have up to as many as 150). But still, deep neural networks are brittle, inefficient, and myopic compared to that of an actual human brain.

Brittle

Deep neural networks can be tricked easily with slight perturbations to the training inputs. A glitch here and deep-learning algorithms start to mislabel objects and other absurd combinations wildly. This critical distinction between biological and artificial neural networks poses a far-reaching challenge for deep neural networks in areas comparable to clinical medicine and autonomous driving.

Inefficient

Data-hungry deep neural networks are inefficient, requiring vast amounts of training examples. Also, how these models are processing them frequently remains a mystery. We can tell a dog from a zebra right away. The brain-like network, however, needs training data to achieve this. This comes to show that human-level insights require the capacity to go past information and deep-learning calculations. People can construct models of the world as they see it, including ordinary common-sense information and everyday common-sense knowledge, and subsequently utilize these models to explain their actions and decisions.

Myopic

Let’s face it. Deep-learning models are strangely intolerant, lacking the same level of cognition. In other words, whereas a human can instinctively tell that a cloud that might have the shape and features of a dog is not a genuine dog, a deep-learning algorithm will have trouble separating between appearing like something and being that thing.

But yet, we call it neural network. If this is not brilliant marketing, then I don’t know what it is. A three-month-old baby has a better grasp of what to make of its surroundings than any deep-learning application built to date.

This Rhetoric Is Misleading at Best and Downright Dangerous at Worst

Humans have always wanted to create machines that can think, learn, and reason. Current research in AI pushes us to look at specific algorithms claiming to be comparable to our human ways of thinking and, subsequently, reasoning. Because of this rhetoric, everybody expects intelligent androids to appear any day. And quite frankly, papers have shown that it’s not only the general public that is torn between science fiction, make-believe, and what can be accomplished.

How can we tell whether to take these descriptions literally or metaphorically?

Here is where it gets tricky. On the one hand, using anthropomorphic tendencies to describe AI phenomena can benefit future research in the field. However, it is also very hindering if not even dangerous in socially sensitive applications. Why? Because the anthropomorphic tendency in AI is not ethically neutral.

What happens when we let algorithms decide in socially sensitive applications? For one, while depending on the data fed into the system, we are potentially faced with racist, sexist, and discriminating outcomes. Secondly, how can we sustain our ability to hold influential individuals and groups accountable for their technologically mediated actions?

It is of paramount importance to understand that the notion of machine learning technologies being humanlike when it comes to their ability to fully understand data (meaning finding patterns and exploiting them) is not correct. While these applications are powerful (e.g. the Optometrist Algorithm), they merely mimic human intelligence.

And that is what is essential here: Such systems are powerful tools for good or for evil. Or, as David Watson put it in a recent article, “The choice, as ever, is ours.”