Neurons are one of the most remarkable inventions of evolution. Modern neuroscience is allowing us to understand how neurons and the brain work at increasingly deep levels. This understanding has led an appreciation of how optimized neurons are for information processing. Some of the brain neatest tricks are being copied by technologists (e.g. self-organization) with more to come (e.g. information and energy efficiency)
Anatomy of a Neuron
We know since the early drawings of Santiago Ramon y Cajal thanks to the neuron staining techniques of Camillo Golgi (shared Nobel Prize 1904) that all neurons share four key elements. The dendrites are the input, they listen to the electric signals of other neurons or to the outside world. The axon is the output, it transmits the electrical signal from the neuron to other neurons or to muscles. The cell body (soma) is the power supply, supplying energy and chemicals for the neuron to function. The synapse is the connection, where an axon transmits the electrical signal to the dendrites of another neuron.
Dendrites are the input terminals, like the keyboard, mouse or internet of a computer. Most are internal, processing the signals from other neurons. A few connect directly to the outside world, allowing us to sense it. Our body has dendrites that are sensitive to a wide variety of stimulus including light (eyes), vibration, sound and movement (ears), physical contact and heat (skin), stretch and force (muscles), chemicals (nose and tongue). Neurons can have thousands of dendrites connected to other neurons and combine those signals with different weights and time delays. The signal processing is done analogically and is very versatile, making scientists think that a lot of our learning could be there.
The axon is the output terminal, it takes the electrical signal the neuron calculates and transmits it to the neurons that are listening to it. An axon is digital, it either has an action potential or it doesn’t. The action potential is a voltage spike from -70mV to 40mV that lasts milliseconds and has a refractory period of ~10ms. The initial segment of the axon is where the input of the dendrites determines whether the action potential happens or not by an extremely ingenious electro-chemical process that was characterized by Huxley and Hodgkin (Nobel Prize 1963). Consequently the axon “speaks” to the dendrites through trains of electrical pulses in the hundreds of Hz in frequency.
The cell body or “soma” is the power supply. Like any other cell, it has the instructions (DNA), the power supply (mitochondria) and the protein factories. It supplies axons and dendrites with the significant amount of energy and chemicals they need to function and keeps the inside of the cell in the right working order to be able to generate and receive the pulses.
The synapses are responsible for communication, they are chemical junctions between neurons that transmit instructions from axon to dendrites whenever the axon fires an axon potential. The instructions are mainly about transmitting the electric signal, but they also include many others that determine if the synapse strengthens or disappears. The synapses are in a way the “software” of the neuron as they determine the strength and type of communication the dendrites of a neuron receive.
Advantages of neurons over electrical circuits
Contrary to what might be thought, neurons are much better at many processing tasks than the computers that we have built. There are three main advantages, all of which technology companies are trying to copy to make computing more powerful
Self-organization and learning.
The first advantage has to do with the plasticity of neurons and the brain. Brain development is extraordinarily complex but it is based on the less than 1 Gigabyte of information humans have in their DNA. The rest is self-organization based on external information and adaptation during development. For comparison purpose, Windows 98 was the last operating system that would have fit in the human genome, with Windows 10 being already 20x the size and any of the current “intelligent” systems like Watson being many orders of magnitude larger.
The self-organization and learning advantage of the brain makes great sense in the context of evolution. Without any explicit design or ample instruction storage capacity whatever emerged had to be very sparsely designed.
Self-organization and learning are starting to be copied which has led to the boom of artificial intelligence around machine learning and deep learning which uses self-organization and learning principles derived from neurons by creating artificial “neural networks”. These techniques are still in the early days with the most complex networks currently in use being probably in the thousands of neurons compared to the brain’s 100 billion neurons.
A second related advantage is the much lower amount of information neurons require to learn. Human and animal brains and neurons are able to learn very quickly compared to machine learning models. A human needs just limited experience with words or driving to perform very accurately, while computers need to “drive” millions of miles or go through billions of words, something which a human would be incapable of doing. The root cause of this advantage is still only speculated about, but it could be around the fact that neurons integrate computing, communication, and information storage together without separating between data, transmission, and computation explicitly.
Information efficiency is also deeply necessary for evolution and highly selected for it. The organism that requires can learn to identify a threat or an opportunity quicker will have a definitive advantage over slower learners.
There is frantic research in getting more information-efficient machine learning models, both in terms of software (e.g. new techniques like deep learning) or hardware (e.g. IBMs neuromorphic chips).
Finally, brains also have an incredible advantage in terms of energy efficiency. Our brain functions with approximately 20Wh per day, this is 20% of the total energy budget of a human that is around 100Wh. However, it is extremely energy efficient compared to a computer with comparable power. According to Forbes a smartphone clocks 1kWh per year, or ~2Wh per day, a laptop according to some sources is around 200Wh per day (so 10x more than a human). If we take the laptop as a reference a supercomputer with brain-like processing power would be using millions of times more energy than a human brain.
Energy efficiency is another cardinal design principle of nature. Brains are already very expensive energetically at a biological level, so the brain has optimized itself as much as possible while keeping its processing power.
Energy efficiency for computers is a key design principle, especially for mobile phones and the internet of things. Apparently combining analog and digital processing like the brain does, could be a key to increased energy efficiency in our own digital tools
Long-term electrical advantages
However, electrical digital computers have other advantages over neurons that will probably allow them to dominate long term.
First, electrical circuits currently work in 10s of GHz of clock speed. This is 10 million times faster than what neurons can muster with their 5ms action potentials and 10ms refractory periods. 10 million times difference is equivalent to the difference between evolutionary time scales (i.e. how long does DNA take to evolve new species) and the speed at which we live our lives and improve our economy and technology.
Second, digital storage allows for perfect recall of data. Our brains are not optimized for exact data, which is quite useless in real life. So computers have an advantage in terms of storing, retrieving and processing detailed information, while the brain is quite adept at extracting patterns and getting the gist of an issue.
Finally, digital computers are almost infinitely scalable. Our brains are famously limited by the breadth of the birth canal, with our heads being as big as they can biologically get. Computers, on the other hand, are being stacked in greater and greater numbers through cloud technologies with potentially limitless processing power and storage.
Will Moore’s law take electronics beyond what neurons can do? It is a distinct possibility, and it could bring an upheaval in how the world is organized. Neuron-only organisms are already at a clear disadvantage against digitally-enabled ones. If that enablement is made more direct through brain-machine interfaces we could have a race of super cyborgs like Yuval Noah Harari describes in Homo Deus. If alternatively, we manage to manufacture conscience in electronics, it might decide it has no use for the outdated neuron-based monkeys that brought it here.
In any case, the power of electronics and neurons has to be harnessed towards our values and goals. Ethics, human rights, global development, freedom and the pursuit of happiness have to be front and center in what we do with our neurons and electronics.