Principles of Neural Design

Chapter 2: why an animal needs a brain.

• Brains allow for control and adaptation within a single lifetime. They make more sense for bigger, longer-lived creatures.
• However, we see adaptive behavior even in brainless bacteria like E. Coli.
  • LACTOSE AND NOT GLUCOSE: E. Coli implements chemical logic to switch between digestive modes. It likes to run on glucose, but can also metabolize lactose. In the latter case it needs special proteins to allow lactose into the cell and then to metabolize it; it prefers not to make these if it doesn't have to. So it has a simple chemical switch: a 'repressor protein' that binds to the relevant genes in the DNA and prevents their expression. The same protein is a lactose sensor: binding to (allo)lactose causes it to detach from the DNA. So the presence of lactose directly activates the mechanism that allows lactose metabolism. At the same time, the presence of glucose suppresses the action of RNA polymerase, which prevents gene expression, so the cell doesn't waste energy on lactose metabolism when both glucose and lactose are present.
  • BIASED RANDOM WALK: the cell moves randomly, exploring, but will hold a particular direction when it senses that the gradient of nutrients is increasing. How does it know this? When a receptor protein is activated, it desensitizes itself by adding a methyl group, so that successive activations mean that the concentration is increasing.
• A paramecium also uses a biased random walk strategy to explore its environment. Because it is ~300000x larger in mass (and 100x as long) than E. Coli, it can move through fluid much more readily, and can traverse longer distances. For the same reason, however, chemical signaling is no longer fast enough. Molecules diffuse within an e. coli cell in 4ms, but would take 40s to diffuse within a paramecium. So it uses electrical signaling instead. Hitting an obstacle activates a stretch-gated sodium channel, which lets sodium ions into the cell, depolarizing the cell membrane. This depolarization then opens voltage-gated calcium channels, which let calcium ions into the cell, which cascades to open more calcium channels, and the calcium signals the cilia to reverse their swimming motion (the calcium is then quickly pumped back out of the cell).
  • Why is this faster than diffusion? If I let in a sodium ion, doesn't it still have to diffuse to near the calcium channel gate before the gate will open? And the same for any new calcium ions. Maybe you could say it's just a matter of distances? That is, the same square-cube law is at play, but electromagnetic forces operate at much larger distances than the nuclear (?) forces involved in chemical binding, so the constants are much lower. Another way to say this is that electrical signaling benefits from speed-of-light transmission over much more of the required distance.
• The multicellular nematode worm c.elegans is itself 2000x as long as a paramecium, and has a real nervous system with 302 neurons.
  • It moves by oscillating its body in a sinusoidal wave. But there is no oscillatory neuron. Instead the oscillation comes from an embodied feedback loop: when a muscle contracts, internal pressure increases, which inhibits motor neurons, allowing the body to spring back into its original configuration. This is much cheaper than computing a sine wave 'in software', and also automatically adapts the speed of the oscillation to the viscosity of the ambient fluid. This is an example of fundamentally embodied cognition: a computation that relies on the physical feedback of the body, so it couldn't be performed (at least not in the same way) by a 'bar in a jar'.
  • Unlike larger organisms, individual neurons are 'identified' with particular functions and have their own individualized sensors, locations, and connections. Computations that might take place across multiple neurons in a larger organism happen using chemical signaling within individual cells. Chemical signaling is much cheaper than electrical signaling.
  • The worm avoids expensive redundancy in sensors and connections by being slow. Instead of reducing noise by averaging multiple sensors, it averages a single sensor over time. It can take minutes for an olfactory stimulus to propagate into a motor signal.
  • The worm does not possess the gene for neural signaling through spikes (specifically a volatage-gated sodium channel). Over short distances, it's cheaper and richer to use fully analogue signaling.
  • The chapter describes the use of 'neuromodulators' (which I take to be the same as neurotransmitters?) octopamine, serotonin, and dopamine for chemical signaling between cells. The amazing thing about these is that they are broadcast widely but act only on the specific neurons that express a receptor. It's like being able to transmit one of many different frequencies, and each neuron chooses which to listen to.

Chapter 3: why a bigger brain?

• This chapter is short and not terribly deep. Larger animals cover vastly larger distances, are exposed to many more opportunities and threats. We need to recognize these and coordinate our motor behaviors accordingly. This involves not just reacting to circumstances but learning to predict future circumstances and rewards.
• There are some patterns, structures, and principles that, although we cannot prove they are necessary, are at least common between fly brains, mouse brains, and human brains, so they seem to be conserved.
• We have an autonomic nervous system, which regulates core bodily functions like breathing, heartbeats, digestion, thermoregulation (sweating, shivering, etc.), and general homeostasis. It is never fully autonomic because it is connected to and influenced by the rest of the brain. We can't directly will our heartrate to increase, but we can get the same effect indirectly by imagining a stressful scene. This is because effective regulation requires prediction, and prediction in the fullest sense requires higher cognition.
• Many constraints on brain design can be seen through the lens of information theory. If a spiking neuron spikes on average R times per second with a timing precision of $\Delta t = 1/T$, then it can transmit $\log T! - \log R! - \log (T - R)!$ bits per second of information.
  • A key assumption here is that $T \gg R$: the neuron's firing rate is constrained so it can't just choose independently whether to fire in any given window. (if it could, there would just be $2^T$ possible second-long spike trains, so we could transmit $T$ bits per second).
  • In fact, spikes are more surprising (carry more information) when they happen less often. Since a spike requires a certain amount of energy, we want to make spikes rare in order to maximize bits per ATP molecule.
  • Faster spiking also requires a linear increase in the diameter of axons, which has quadratic cost in space.
  • Both of these incentivize us to send at the lowest acceptable rate. Which means we must send only what is needed.
• We also want to minimize wire, because wire costs space and energy, but also because neural signaling is relatively slow, so minimizing processing time requires minimizing distance traveled.

Chapter 4: how bigger brains are organized.

• Local connections are segregated from long-distance connections. In mammals, axons of more than a few millimeters are myelinated, that is, covered in myelin. These are called white matter.
• The daily cycle of activity and inactivity (wake/sleep, roughly) corresponds to periods of catabolism (breaking things down, e.g., fat->fatty acids->glucose) and anabolism (building things up). These often involve running the same reaction in reverse, so you don't want to do both at the same time.