4
$\begingroup$

This question assumes a definition of AI based on machine learning, and was inspired by this fun Technology Review post:

enter image description here

SOURCE: Is this AI? We drew you a flowchart to work it out (Karen Hao, MIT Technology Review)

As the definition of artificial intelligence has been a continual subject of discussion on this stack, I wanted to bring it to the community for perspectives.

The formal question here is:

  • Is an algorithm that is no longer actively learning an AI?

Specifically, can applied algorithms that are not actively learning be said to reason?

$\endgroup$
  • $\begingroup$ PS: I'm claiming fair use here on the image--link for the article directly below. $\endgroup$ – DukeZhou Nov 22 '18 at 0:27
  • 1
    $\begingroup$ Interesting question, but this boils down to the definition of AI. $\endgroup$ – nbro Nov 22 '18 at 0:36
  • $\begingroup$ @nbro it does, but asks specifically if learning is the requirement. For instance, is an NN that has been trained but is no longer actively learning an AI? (Interesting DeepMind post on Enabling Continual Learning in Neural Networks.) $\endgroup$ – DukeZhou Nov 22 '18 at 1:01
  • 1
    $\begingroup$ A usual there needs to be lines drawn between "classic AI", maybe "classic CI (computational intelligence)", "modern AI in practice", the (very vague) AGI. An AGI that could not learn would be severely hobbled - Human: I am Bob. AGI: Sorry, I forgot, what was your name? - unless you want to include memory but exclude learning, and the possible difference between the two would be an interesting question in its own right. $\endgroup$ – Neil Slater Nov 22 '18 at 9:36
  • 1
    $\begingroup$ @NeilSlater that's a very good point: a computer game that asks for the player's name, is that "learning"? It doesn't have to be an AI... I think that line is going to be a very fuzzy one! $\endgroup$ – Oliver Mason Nov 22 '18 at 10:02
2
$\begingroup$

In my view learning is a useful aspect of an AI system, which makes it more resilient towards a changing environment, but it is not a requirement. Early AI systems (Expert systems, theorem solvers) would usually not use learning algorithms, but had a fixed knowledge base.

It also works a lot better with numerical models than with symbolic processing; the proliferation of training data which boosts the most recent AI wave (which is based more on statistical and numerical methods) has made it somewhat necessary for modern systems to have a learning/training component.

However, one needs to distinguish between the initial training, which leads to a production system which then becomes fixed (and will not learn from new input) and those which have a learning component built in as an intrinsic part and will continue to adapt throughout their lifetime.

A cautionary example is Microsoft's "Tay" chatbot. This was still actively learning from conversations and was quickly abused by subjecting it to undesirable input. Switching off the learning component would have been prudent in this case, but it would not have taken the AI out of it.

So, to answer the question: learning is not a required part of an AI system once it is operational. While modern systems generally have some element of learning to build their initial models, older systems were simply hand-coded without any automated learning at all. But one would still regard them as AI system.

$\endgroup$
  • $\begingroup$ It would be great if people who downvote answers would leave a comment explaining their reasoning. $\endgroup$ – Oliver Mason Dec 3 '18 at 10:51
0
$\begingroup$

Is an algorithm that is no longer actively learning an AI?

This question unfolds into two, where the term tuned means configured with parameters that were derived through some optimization process.

What segment of the AI community thinks an algorithm developed or tuned through AI techniques is, by itself, an intelligent component?

Should the AI community classify an algorithm developed or tuned through AI techniques as an intelligent component?

To meaningfully address the first question, one would need to poll the community in a way that generates a representative data set and analyze the data to produce a reliable response. These are the likely factors that would cause a person to lean toward stating that an unchanging algorithm with unchanging parameterization is cannot, by itself, be considered an intelligent component.

  • Genetic algorithm researchers
  • Geneticists
  • Neuroscientists
  • Electrical engineers
  • Statisticians

The answer to the second question is, "Unknown at this time and possibly impossible to know." Here's a story to explain why.

Polly reads about the tie between sugar and diabetes and because her mother has to take insulin, she learns to read the labels in the grocery store and avoid holiday candy. Her friends and family think she is intelligent. She has the neural equivalent of a running algorithm so that she can recognize brands and products that are sugar free.

The food industry, following administrative guidelines literally and considering possible legal risks, stop adding cane sugar to many products. They discover that high fructose corn syrup produces an improved behavioral response, specifically, higher purchase rates, for ingredient concentrations that correspond to an equal expense in production costs between the two sweeteners. The boards of the food corporations find that to be intelligent.

Polly is diagnosed with diabetes. She later discovers the substitution, changes her shopping habits and when her friends and family ask her what happened, she drops her eyes and mutters, "I was an idiot to trust corporations."

Was Polly smart to develop the sugar algorithm and then stupid to not modify it to include all sweeteners that convert to glucose in the blood? Was applying a financial algorithm to the selection of ingredients smart or simply a mechanical momentum in the space of information theory?

There are several interactions in this story that illuminate the difference between discrete beliefs, language, prediction, risk analysis, and the class of algorithms that could deal effectively with those mental constructs. Whether an algorithm exists that is generally intelligent such that it will develop whatever abstractions are necessary and process behavior ideally in the pursuit of corporate or medical health or any arbitrary set of objectives is not yet known.

We have three more questions that are central to the line of inquiry represented by the flow chart in question.

Is the mind a meat machine (Minsky) that will eventually be cracked and replicated in VLSI design?

This question once considered philosophic has transitioned clearly into the realm of the scientific and mathematical. There may be an example created proving the affirmative beyond any reasonable doubt. There may also be a mathematical approach to prove the affirmative or prove it as theoretically impossible.

How many lines of Java, Python, or whatever general purpose imperative programming language is necessary to realize all the capabilities of the mind in a single algorithm or class of them, delegating all the particulars of domain, strategy, and every other aspect of what we think of as intelligence as mere data that drives the processing of input to output by this single unchanging algorithm or algorithm class?

The answer to this question is either a number found by counting a working system based on a developed theory and actual demonstration or a variable for which there is no known value at some particular point in the timeline of AI research. If this second question is answered with a number, then we have at least one fixed algorithm that is intelligent even though its signal paths are entirely driven by fixed parameters. Only the data travelling through the signal paths are variable.

This, incidentally, exceeds the brain, since its paths are neuromorphic.

If the immediately above question be answered with a number derived by a proof of concept, is it still anthropomorphic, working effectively in its ability to match or exceed human mental abilities but not a general problem solvers that will work effectively anywhere in the universe on any problem, whether or not humans can imagine it, and find some optimal solution, determining dynamically what optimal means in every possible context?

This question can be decimated by logical induction applied to any one of the absolutes. At no point in time can a human brain or any combination of them determine with absolute certainty any of the following things.

  • Every kind of problem that can exist in the universe
  • Every kind of environmental scenario in the universe
  • Every kind of objective for which optimality has meaning

This last limitation can only be overcome if a single set of equations or some kind of future way to model the universe and some means of measuring everything in the universe is overcome to test the model in every possible location and over a substantial length of time. Most would consider that level of optimism so extremely humanistic that it qualifies as meglomania.

This last reality is why intelligence is inescapably tied to context.

$\endgroup$
-1
$\begingroup$

The classical definition of AI goes into the direction of a controller. A so called PID controller is a piece of code which is able to drive a robot on a line and adjusts the wheel position. After pressing “go”, the PID controller aka AI is doing something. In reinforcement learning the idea is to extend the simple controller into a self-modifying controller which can learn to follow the trajectory better, for example by adjusting it's internal q-table.

The problem with this strict definition of AI as a controller is, that it ignores authoring tools which are not controllers but tools for creating controllers. An authoring tool is for example a scripting language. The scripting language itself is not able to control a robot on a black line, but it provides commands for building a controller. The reason is, that scripting languages provides commands like “left, right, detect-line” and so on and with this building blocks it is possible to create a macro. Summary: There are many examples of non-learning AI systems which have to do with Artificial Intelligence but can't be categorized as a classical controller.

$\endgroup$
  • $\begingroup$ "The classical definition of AI goes into the direction of a controller" where do you find such definition ? You wrote an answer based on an example... that's no how it's work. You use example to illustrate your answer, not to answer $\endgroup$ – Jérémy Blain Nov 22 '18 at 9:05
  • $\begingroup$ @JérémyBlain Robotics programming can be called classical AI application. Most people who are interested in programming intelligent machines are starting with a line following robot. And the recommended first task in programming a robot is the line following task which is realized with a PID controller. A possible argument against this definition is to reduce AI to a philosophical debate in which concrete software is offtopic. $\endgroup$ – Manuel Rodriguez Nov 22 '18 at 9:11
  • $\begingroup$ There is much more to AI than robotics... $\endgroup$ – Oliver Mason Nov 22 '18 at 9:24
  • $\begingroup$ I embrace reductionist views, since my own view of "intelligence" applies to the most reduced forms of decision making (intelligence as a spectrum and dependent on context.) It also seems that the term "AI" is used informally, so a simple tic-tac-toe algorithm that utilizes a solution (as opposed to learning) is referred to as an "AI". Similarly, before the term was popularized, people might have said "the computer" controls the process in regard to a control system. $\endgroup$ – DukeZhou Nov 25 '18 at 22:00

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.