Stop using biological analogies to describe AI. It’s 99.999% wrong.
Before I transitioned to machine learning and data science, I spent many years doing research on biological topics. Probably that’s what always gave me a light shiver whenever someone introduced deep learning with biological analogies. Many of these analogies are somewhere between wrong and misleading and can be found in a lot of the common AI terminology (such as the term “A.I.” itself). I long saw this as something unfortunate which must be accepted and dealt with. But I recently changed my opinion. Now I believe that we (as people working with “AI”) need to make a much bigger effort to change the narrative, the way we communicate about “AI”.
Why I changed my mind?
For a very long time, I underestimated the negative effect of poor terminologies and questionable (or plain wrong) biological analogies. But I slowly came to the conclusion that it actually did a lot of harm to the outside perception of the field of “AI”, and introduced severe cognitive biases among “AI” researchers, experts, practitioners. So, let’s work together on changing this!
I will first discuss biological analogies and then terminology.
Biological analogies
It is an extremely common slide in instructions to deep learning. It is in many books on deep learning and it is on Wikipedia: A visual analogy that depicts a sketch of a biological neuron and its artificial counterpart. It makes a good story and there is a historical connection in form of inspiration (bionics, https://en.wikipedia.org/wiki/Bionics), so why not?
Let’s take an entirely different example: planes. It seems OK (to me) to have a bird sketch next to a plane picture to show where the inspiration came from. So why not show a biological neuron to illustrate how artificial neural networks were invented?
Simply because the level of analogy is a very different one. The wing of a plane uses the same physics to generate the same type of uplift a wing of a bird does (doesn’t flap though, I know, …). An “artificial neuron”, however, is hardly mimicking biological neurons at all. The only analogy really is that some kind of signal goes in, is processed and then, maybe, some other kind of signal goes out. I find that far too little of a relationship to have real explanatory value.
I bet most of us would find it ridiculous to say “artificial bird” when we speak of planes, even though the underlying analogy is technically much closer. Unfortunately, talking of “artificial neurons” is common AI-speak though. One reason why we cringe less when speaking about artificial neurons compared to artificial birds might be that everyone is very used to birds, but few people really know much about neurons.
An artificial neuron multiplies incoming numbers and applies a function to the sum of it (the “activation function”). Biological cells in contrast, are incredibly complex entities able to do extensive signal processing themselves (see the wonderful illustration above by David S. Goodsell). A single biological cell is so complex that it needs many, many thousands of researchers to work on understanding it a bit better, yet none of them would argue that we fully understand how it works. On top of that, neurons form networks that in most cases look entirely different from artificial neural networks. Neurons in our brain are not forming a linear stack of layers. And, no, we also don’t learn by backpropagation! An artificial neural network is hence nowhere near a network of biological neurons! The omnipresent neuron comparison slide should hence be banned and only be allowed in extensive sections on the history of deep learning. Please, don’t make it the cover picture or main story line.
The main reason why this analogy troubles me is not because it is wrong from the biology side of things (it is wrong though). I think the main harm of this analogy is that it drastically misleads. Funny enough, this holds not only for a lay audience, but also for many people working in the field of “AI”!
Experts tricked by their own over-simplification
A few years back there was a large debate in the field about a prognosed event termed “singularity”, which refers to the point in time when “AI is taking over”, or in a little less dramatic: where computers surpass human intelligence in a general sense and will only become more and more intelligent. Those speculations where also driven, or at least partly believed and supported, by a lot of people that actually had substantial inside into how algorithms and machine learning techniques work. Prognoses on future capabilities of artificial intelligence have a long history of dramatically wrong predictions. In that sense it was no complete surprise. Still, it gained large media attention and remains an obligatory question at many public events on “AI”. In the case of the “singularity” the sudden fear even among tech-people was clearly driven by a long streak of unexpectedly large gains in performance across many areas (first computer vision, then later natural language processing). But it was also supported by many false analogies between AI and biology. My favorite examples are frequently done comparison between the number of artificial neurons in current and future deep learning networks with the number of neurons in a human brain. Or comparisons between the weighs of a deep learning networks (“the learned numbers”) and the number of synapses. By such measures we are very rapidly approaching human-like levels*. And yet we are no step closer to a general artificial intelligence then we were 40 years ago. In fact, many insects have better capabilities in moving through the real world than any robot we ever built even modern robots can easily have orders of magnitude more “neurons”.
You can check a list of animals sorted by number of neurons and you will see that we long operate deep learning networks which should far exceed the capabilities of many species if we make the mistake to believe that the number of neurons is a relevant measure. https://en.wikipedia.org/wiki/List_of_animals_by_number_of_neurons
The above gif is to show how wrong the analogy “artificial neuron — neuron” is. It shows a single living cell (left, a macrophage chasing bacteria) which is perfectly capable of moving through its environment in an “intelligent” way, while a robot (right, probably having 1000s of “artificial neurons”) clearly is not.
If the number of artificial neurons would be a meaningful measure of intelligence, than why can even insects navigate real world better than any of our robots? If at all, the number of nodes in a network is a measure of fitting capability. But better fitting functions do not automatically mean “higher intelligence”. If used properly, with the right data at hand, it can mean higher ability to mimic intelligent behavior, but that is something else entirely.
The wrong biological analogy has hence not only misled media or science fiction authors. It has misled many prominent figures from tech-land, too. Maybe because it is such a catchy story and makes it easy to get a false sense of intuition. In my view this is also a very strong case for more inter- (and multi-) disciplinary research because I doubt that a cell biologist would have fallen into the same trap. Better don’t fully leave it to tech-engineers with only a cartoon-sketch-level of understanding of a biological neuron.
Poor terminology.
Many terms that we (we as the machine-learning/AI community) use are, at the very least, unfortunate choices. Unfortunate, because they are highly misleading. Terms such as “intelligence”, “learning” or “neural network” trigger expectations that in the best case are overly optimistic. In the worst case they are as far from reality as it can be.
Why the terminology matters.
An example. If I would tell you “I just bought a nice house”, you might immediately feel envious, astonished, or you might want to congratulate me. But I meant a Lego house, 15€, I just bought it for my daughter. Technically, I used the word “house” correctly, yet given the context it must be considered a poor choice of wording. “Artificial intelligence” has as much to do with what most people intuitively understand under “intelligence” as my Lego house with a “real” house. One might argue that the terms are still totally fine, because after all I speak of “artificial intelligence” and not “intelligence. Isn’t that a proper enough distinction? If I speak of a “Lego house” this indeed does the trick. If I say “I just bought a nice Lego house” using the term “house” no longer misleads. But while “Lego house” is an own term that everyone understands and has a clear picture of, that cannot be said about the term “artificial intelligence”. “Artificial” as the opposite of “natural”? Both are terms that keep slipping away whenever you want to nail them down. We don’t know what artificial actually means, but we believe we get the sense of “intelligence”, so that will dominate our perception when we hear or read it (I believe this could be another case of availability bias).
What to do about this?
I wish I had a good answer. Many researchers already avoid “AI” because it simply feels silly. I would never tell a colleague from the same field that “I used AI for X”. I would always say “I used machine learning for X” or “I used deep learning for Y”. Many colleagues do the same and I would thus consider that those terms are at least a bit better. I also agree with many other researchers that “machine intelligence” would be better than “artificial intelligence” because that triggers associations that tend to be off less. But what about “neural network”? The term is very old and thus extremely established in the field. Won’t be easy to get rid of that and there is no very catchy alternative around that I am aware of. Backpropagation-based function approximators? Rather not, I fear.
Do you have any suggestions?
Another way out.
What if we won’t change our terminology? Well, maybe we are too late for that anyway. In my own perception the connotation of “artificial intelligence” is already changing fast. Unlike a few years ago, the silly debates about a “singularity” luckily are no longer very loud. And the more people get used to “AI” in everyday life situations the more ironic the undertones get when people say “AI”. When I talk to colleagues, and someone uses the word “AI” it is usually pronounced with a very strong irony intonation:
Hey, I have this new project related to <strong irony>”A.I.”</strong irony>…
The silly this sounds, I try to take this as a good sign showing that it eventually will go without saying that “artificial intelligence” has virtually no resemblance with what we usually mean when we speak of intelligence in our everyday lifes.
Nothing new, I am only rephrasing.
The main topic is not new. Others have complained about poor terminology long before me, e.g. https://towardsdatascience.com/deep-learning-versus-biological-neurons-floating-point-numbers-spikes-and-neurotransmitters-6eebfa3390e9
Related references worth reading:
- “The application of artificial neural networks in metabolomics:
a historical perspectiv”, Mendez, Broadhurst, Reinke, 2019, https://t.co/Kki4vRxI30 - Other arguments for why A.I. is not a great tag if we actually mean machine learning: https://www.theguardian.com/commentisfree/2022/nov/05/machine-learning-systems-are-problematic-thats-why-tech-bosses-call-them-ai?CMP=Share_iOSApp_Other
*I say at one point that we rapidly approach human-like numbers if we would — wrongly- compare real biological things to the “artificial” version. I was refering to a common comparison of model parameters vs. synapses. Some of the biggest models to date are the new, huge language models, some with more than 10¹² parameters (https://dl.acm.org/doi/pdf/10.1145/3442188.3445922) vs. 10¹⁴ synapses in the human brain… But again, that comparison makes no sense at all. Comparing apples and oranges is far better than this.
