In 1968 France entered into a social eruption, an occurrence she is very used to. At the same time, she launched on the airwaves of the ORTF, a television series that was as original as it was absurd: the Shadoks (Rouxel, 1968).
SwissCognitive Guest Blogger: François Blayo, chief scientist officer NeoInstinct
The Shadoks were “anthropomorphic creatures with the appearance of chubby birds, with long, filiform legs, tiny and prehensile wings, and original hair.” (Wikipedia). Living on a planet with uncertain contours, their main life goal was to build a rocket to land on the earth. To achieve this, they invented the “Cosmopump” intended to pump the “Cosmogol 999” to fuel their rocket. And so, the Shadoks pumped and pumped and the French made fun of these poor Shadoks…
Let’s forget the Shadoks for a moment and focus on the effort of reading that led us to these lines. Our brains consumed energy during this turmoil. But deep down, what do we know about the energy consumed by a brain
in action? Our brains represent about 2% of our bodyweight but burns 20% of the nutrient energy and oxygen that we absorb. 80% of this energy is consumed by neurons, and the remaining 20% by glial cells. The majority of this energy is consumed at the synapses level, those minuscule spaces between the neurons where signals are sent and received. There, neurons constantly pump ions into the space that separates them by exchanging potassium and sodium to create electrical charges. This pumping action is fundamental for the operation of brain circuits, but it is very energy-intensive.
Our neurons function like the Shadoks: they pump permanently.
Does the brain consume this energy uniformly? Definitely not. Our auditory process mobilizes more energy than the olfactory system or the areas of the brain responsible for memory. Hearing requires rapid and accurate signalling, while the sense of smell does not have such intense energy needs.
Does the brain consume energy during sleep? Yes, clearly and even in an almost identical quantity to that consumed when awake. The brain never stops. Neurons maintain an active communication during sleep,
strengthens synaptic connections, and consolidates what it acquired. And what happens when we focus on a particularly engaging reading or a difficult statistics problem? Curiously, energy consumption increases little in relation to its nominal consumption. Only by a few percent.
And what about the statement notably staged in the film ‘Lucy’ where human beings only use 10% of their
brain’s capabilities? In reality, neurons are mostly silent for long periods of time before coming into action when it is necessary. For example, speaking: there is no need to consume energy when you are not speaking. But being silent does not mean that the neurons are not activatable. They can be at any time. Rest assured, we use 100% of the neurons in our brain (Rodo, 2018). There is no need to look for a miracle product to achieve this Nevertheless, what is remarkable is that our brain consumes really very little energy: on average a brain consumes 20Wh.
Let’s examine the energy consumption required by running algorithms. We can quickly state that something is wrong. For example, learning and optimizing the architecture of the Transformer model (Vaswani, et al., 2017) recognized as one of the best models for machine translation, can emit five times more CO2, equivalent to a car during its entire lifetime, including the energy needed to build it (Strubell, Ganesh, & Mc Callum, 2019). This is equivalent to 656,347 kWh of energy consumption to obtain this machine translation model, or 3,746 years of a human brain’s consumption!
Without being very good at languages, many humans are able to learn one or even several in less time than that and by consuming far less energy. If intellectual economy offered by for the programming of these translation tools is real, it is realized at the cost of an unacceptable ecological footprint. This obviously raises the question of the algorithms used. It is clearly not possible to continue developing such energy-inefficient . So, there is obviously a lot to review before starting to talk about intelligence. The Shadoks, in addition to pumping on a daily basis, constantly refer to principles and proverbs of their own. One of which says: “by continuously trying, we end up succeeding… So, the more we fail, the more likely that it will work.”
Shouldn’t we act like Shadoks with today’s Artificial Intelligence?
About the author:
François Blayo is an Artificial Intelligence pioneer, expert, author, speaker and popularizer, currently leading research at NeoInstinct.
Jolivet, R., Coggan, J., Allaman, I., & Magistretti, P. (2015). Multi-timescale Modeling of Activity-Dependent
Metabolic Coupling in the Neuron-Glia-Vasculature Ensemble. PLOS Computational Biology(26).
Rodo, C. (2018, January 17). At the borders of the brain. Retrieved from CNRS Le Journal:
Rouxel, J. (Producer). (1968). And here are the Shadoks, season 1 | INA Archive [Film]. Retrieved from https://www.youtube.com/watch?v=tpD0Pdr7oD0
Strubell, E., Ganesh, A., & Mc Callum, A. (2019). Energy and Policy Considerations for Deep Learning in .
Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, (pp. 36453650). New York.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., . . . Poloshukin, I. (2017). Attention is all you need. 31st Neural Information Processing Systems.