THE
WESTERN INSTITUTE
FOR
ADVANCED STUDY
HOW DO OUR BRAINS ENCODE INFORMATION?
Our spinal circuits give commands, but our cortical circuits make predictions.
It is the job of neuroscientists to explain how neural systems do biophysical work in a way that produces conscious experience - and to explain what the necessary and sufficient criteria are for consciousness, so that we know what other systems might generate conscious experience.
​
A good starting point is to compare spinal reflex circuits and cortical neural circuits. When we touch a hot stove, temperature receptors and pain receptors in the skin are activated. The sensory neuron receiving these signals is activated, and it sends a signal to an interneuron in the spinal cord. That interneuron then sends a signal to an alpha motor neuron, which fires. That impulse causes the ipsilateral muscle to flex, thereby achieving the limb withdrawal response with only a simple three-neuron circuit! But there is no perception of the pain, until a second or two later, when the information reaches thalamocortical circuits in the brain. And of course, there is no decision to withdraw the limb from the hot stove – this is a spinal reflex, after all. So there is something about perceptual experience and the ability to decide our actions that is deeply tied to cortical neural network activity, and is not achieved by just any neural circuit.
​
So what is it about cortical neural circuits that is special, compared with spinal reflex circuits? There is a critical difference, one we have known in neuroscience for about thirty years. Spinal neurons have essentially deterministic outcomes – they are always in an off-state or an on-state, firing or not firing, at any given time. By contrast, cortical neurons have probabilistic outcomes – they are essentially calculating the probability of transitioning from an off-state to an on-state. Cortical neurons tend to sit right at the threshold for firing an action potential, and they allow random electrical noise (that is, stochastic events) to affect signaling outcomes. As a result, individual cortical neurons have statistically random inter-spike intervals. And at the network level, a statistically random ensemble of cortical neurons fires synchronously, and then a few milliseconds later, another statistically random ensemble of cortical neurons fires synchronously. These so-called ‘cortical oscillations’ are thought to be critical for conscious awareness.
​And of course, when we observe order arising from randomness, we do tend to think of quantum systems. After all, a modern quantum computer looks just like that! The qubit has some probability of being in a spin-up or spin-down state, and if left alone, it will collapse into one state and complete a computation. Is that what cortical neurons are doing? Maybe. I believe so. But we need more formal hypotheses to test in the lab, like the decoherence timescales of the system and the observation of quantal energy release in the brain (these are specific hypotheses prompted by this new theoretical framework). Notably, Chris Fields has also been exploring the quantum computing approach as well, to describe cortical neural circuits. Turns out, it's really useful to think about a cortical neuron as a two-state quantum system.
But it's a little more complicated than a two-state system, when you dig into the mechanisms underlying the neural computation. To really understand this system, we have to expand the dimensionality of quantum computation. Instead of spin-only qubits, which have only one axis of uncertainty, we can model each ion as having some distribution of possible positions and momenta, in relation to the neuronal membrane. This probability distribution can be modeled along the x, y, z, and time axes (it's also sensible to include the atomic orbital, which is also uncertain). Instead of having a distribution of possible system states (information) along one axis, we have a whole manifold of probability distributions. That gives the system way more computational power. Imagine each ion holding this much information, and the neuronal membrane encoding all these possible system states within its electrochemical potential! This high-dimensional probability distribution is quantum information. And this high-dimensional wavefunction collapses, as the encoding system interacts with its surrounding environment. We can think of this process as a constructive and destructive interference of the probability amplitudes, across all these axes. As a result, the system completes a computation, and all the neuronal states are defined by how all the ions actually moved around. Another interesting thing happens in this model – since these wavefunctions are constructively and destructively interfering on the polymer surface of each computational unit (and the network as a whole), the process encodes information on these holographic recording surfaces, and naturally generates a holographic projection of all the information encoded by the system. That holographic projection changes over time, and encodes the location and intensity of stimuli from all contributing sensory apparatus. Pretty neat – you get qualia for free, by formally modeling out the high-dimensional quantum computation in a biological circuit! ​
​
So when we observe the world, we are extracting signals from the noise. We are completing a computation – and this is just a best guess. The cortical neural network is changing itself to encode the state of its surrounding environment. It is making a prediction about the state of the surrounding environment – and that prediction is going to be thermodynamically limited by the time and energy resources available. If something is so outside of our expectations, we might not even clock it. So really, this whole framework is really about predictive processing, and it’s a way of describing a practical neural implementation of Bayesian Inference or the Free Energy Principle.