Mark Humphries was always interested in coding foremost, and became interested in neuroscience in college. This book is intended to be for a general audience, and summarizes the last couple of decades of work in systems neuroscience. He also intended it to be for professionals in non-neuroscience fields like clinical psychology and AI development who might benefit from learning about the latest research in systems neuroscience. Systems neuroscience is defined by Humphries as the study of how neurons interact with each other. He noticed that while there are a lot of popular and accessible books of neuroscience, there are almost none for a general audience on systems neuroscience.
Your lifespan is about 34 billion billion cortical spikes. We’re in the midst of the golden age of systems neuroscience. This is mostly because we can now measure electrical impulses from many neurons at once, and we can turn them on and off with light (optogenetics). We can test directly what spikes are for.
In the 1920’s, we were recording our first spikes, tiny little electrical perturbations on an oscilloscope. Based on the fact that they always had the same size and shape, McCullough hypothesized they were all or nothing. This turned out to be the case. Because of this, McCullough realized we can simplify how we think about the brain, since each signal from a neuron is binary. He would work with Pitts to show that neurons sending binary signals can perform all logical computations.
Hodgkin and Huxley in 1952 recorded from giant squid axon. Outside neuron is lots of sodium (positive charge) ad chlorine (negative charge). Inside is a little sodium, a little chlorine, and lots of potassium (positive charge). The inside of the neuron is much more negative than the outside at rest, which creates a voltage across the membrane of the neuron. Hodgkin and Huxley found that when neuron’s voltage increases beyond it’s tipping point, holes open that only permit sodium, so sodium rushes in, voltage goes up briefly. Onrush of sodium triggers the opening of potassium channels, which pumps potassium out. The outrush of potassium shuts off the sodium channels, so voltage becomes negative again. This rapid rise and decrease in voltage is the spike. They made a set of mathematical laws that detailed which holes opened when, and what changes that lead to in voltage.
Von Neuman actually read about the logic computations of neurons and modeled the first computer on that. So the computer is like a brain, not the other way around.
In the eye, most of the initial cells for vision don’t fire spikes. So why do we even need spikes? Spikes though, let neurons send information accurately, fast and far. Spikes timestamp things with extreme precision.
Why do neurons turn voltage into chemical transmission, and then back into voltage? This gives the brain flexibility. Since the action potential is all or nothing, the chemical transmission portion allows the signal to be graded. The more neurotransmitter is released into a synapse, the more likely it is to elicit an action potential in the receiving neuron. But, the creation of a single new spike is the result of a legion of other neurons stimulating it, often a couple of hundred. About 90% of these inputs make it more likely to fire, about 10% are inhibitory, making the neuron less likely to fire. This is essential.
The total amount of excitation and inhibition in the cortex adds up to roughly cancel each other out so neither one gets out of control. This is also true at the level of each individual cortical neuron. Where an input arrives on a neuron can control how big of a voltage spike it will make.
- How close to the body of the neuron is the input?
- How bunched together are the inputs?
- Are there inhibitory neurons between the input and the body?
Simple cells in the visual system detect types of edges. Complex cells combine inputs from simple cells to detect combinations of edges, or certain kinds of contrasts.
Interneurons are inhibitory. They serve to clean up signals. They enhance signal from like-minded neurons, and suppress others. These tend to have a lower failure rate than excitatory neurons.
Pyramidal neurons in the cortex are all excitatory and use glutamate. All of them connect with neurons of the same type within each layer. But they are separated by where they send their axons.
Highway “what” is a bundle of axons that determines what things are.
Highway “do” is a bundle of axons determining what you need to know to do something: distance, location etc. Motion sensors live here from simple directional sensors in V1 to global motion sensors in area MT which integrate simple inputs from V1 and V2.
The failure rate at synapses differs a lot. Some have failure rate of 95%, some of 0%. Failure means an input neuron failing to elicit a change in voltage. This variability is actually a powerful computational tool. Synapses have to start as unreliable connections because the strengthening of a connection is how learning happens in the brain. The failure rate is under the control of the receiving neuron. It can function, for example, to set a threshold for fleeing or staying in the presence of a threat. Interestingly, noise in the brain may also serve to generalize between instances.
Additionally, when decision making, making small adjustments to a solution are often worse than the original solution, but a large jump can yield a much better solution. Synaptic failure thus may be at the heart of the brain’s search algorithm.
The dark neuron problem: Most neurons in the cortex are not firing spikes at any one time. Previously, experimenters could only find neurons by the fact that they were spiking, so we previously didn’t know about this. There is an average of one spike per second for every neuron in the cortex. But 90% of neurons in the cortex are silent during any given one second period. It turns out, about 10% of cortical neurons produce half of all spikes. Why are dark neurons there? Are they waiting to represent things we haven’t learned? Or perhaps they do fire, just not often. Synaptic failure could, paradoxically allow this and amplify neurons that rarely fire. It’s also possible that the situations in which we record from neurons in the laboratory are not very stimulating.
He coins the term Type 2 dark neurons, which represent nothing about the outside world. They don’t respond to changes in the outside world, but they are sending spikes. They’re dark with respect to the outside world. That’s very confusing. I think he just means spontaneous spiking neurons.
A common misconception is that all spikes are caused by events in the world. Most spikes are actually spontaneous. This is especially true during sleep. During development, there is a lot of spontaneous activity which helps to dictate the organization of new neuronal networks. Neurons don’t even need stimulation from another neuron to fire one. They can generate one de novo. An output neuron from the basal ganglia for example, will spontaneously generate rhythmic spikes even in a petri dish. To do this, these neurons are equipped with special ion channels. As voltage plummets at the end of a spike, these channels open, letting positive ions slowly accumulate in the neuron, driving up its voltage again. It’s a feedback loop. It’s the same mechanism as the pacemaker cells in your heart. Pacemakers, similarly, are all over your brain, in all parts of the basal ganglia, the globus pallidus, subthalamic nucleus. A critical group of pacemakers in the midbrain transmit neurotransmitters all over the brain. If the flow of dopamine in the striatum is interrupted, you get Parkinson’s disease.
Spontaneous spikes also seem to play an important role in prediction. For example, in people trying to decide whether a picture is of two faces or a vase, the more spontaneous activity in the fusiform face area before seeing the picture, the more likely they are to decide it’s a face. The spontaneous spikes in these cases, may represent previous experience. Interestingly, as an animal grows older, the differences between spikes evoked from experiencing the world and spontaneous spikes decreases. This is thought to reflect greater learning about the natural world, and predicting it accurately. Information about the outside world may actually be encoded by changes in spontaneous spikes. In order to experience the world quickly, we can’t rely on neurons to generate signals from scratch each time. Experiences making tweaks to ongoing spontaneous activity is much more efficient. The neurons that receive spontaneous inputs are thus nearly ready to fire if they need to.
Spontaneous activity is you. They are your daydreams and planning, your musings. Your rich internal life is the sending and receiving of spontaneous spikes across your brain. “The most important journey for the spike then, is not from input to output, it is the eternal cycle, looping forever within the brain.”
The future: we’re getting better at recording from more and more neurons at once. Companies like Neuralink are working on implanting electrodes into healthy brains, and we could record from them for years.
Discussion points:
Below is a link to a paper on long-term ensemble drifts of representations that Ben brought up in the discussion. This paper is looking at the posterior parietal cortex. The observation of these types of drifts are strange because we usually imagine that stable populations of neurons continuously represent some learned information. It’s weird that over many days, this neural representation moves around. The authors emphasize that neural representations in motor cortex and in primary sensory cortices are extremely stable. Ben brought it up because he thought that potentially these “drifting” neural ensembles may be involved in the silence of many neurons.
https://elifesciences.org/articles/51121
Below is a link to a paper about how ongoing spontaneous activity can influence visual object recognition. Images were presented near the contrast threshold or where a person would be able to tell what category the image was in. Fluctuations in spontaneous activity are by default more predictive of certain categories than others, and so you could predict which objects were identified based on the state of the spontaneous fluctuations. Ben brought this paper up, in part, because Heather was bringing up the important point of how ongoing, spontaneous activity might affect the autonomic nervous system. That made him think of this paper, which suggested that autonomic activity strongly affected the ongoing neural activity. Fluctuations in arousal increased sensitivity to object recognition generally and correlated with pupil size. By the authors of this paper also found that there was a more content-specific effect of ongoing activity that predicted what kinds of objects could be perceived at any given time. I thought that this finding was really cool!
https://www.nature.com/articles/s41467-019-11877-4
The mystery of Nissl staining
There is a type of staining used on neurons, called Nissl staining. When we were in graduate school, we learned that it was a mystery why only about 10% of neurons would get stained. We’re not sure if that mystery has been resolved. Since we learned that so many cells are inactive for any given measurement period, we wondered if this would help explain the Nissl staining mystery. It seemed plausible to us that the stain would only be able to enter active cells as pores open to let in water and ions.