Notes: Principles of neural design – act65com

Skip to content

Search

Search for:

The books starts with some great questions:

Why does an animal need a brain?

Why does it need to be big?

Why does an animal need a brain?

I learned something important here, brains are not for learning, they are for quickly communicating information from sensor to actuator.

One impediment to richer behavior is that there is only one cell membrane and thus only one line for fast (electrical) communication.

Consider a small bacterium that bumps into some obstacle, the collision sensor on its head needs to send info to its flagella on it’s end. This communication can occur quickly because a bacterium is small and thus chemicals have a small distance to diffuse across. But now consider a multi-celled organism in the same setting, collision sensors and actuator. How can information be quickly communicated between locations many cells apart? Electrical communication is a good candidate as it is fast.

Why does it need to be big?

Coordination demands some mechanism with an overview that enables it to weigh alternatives, set priorities, and then exert ultimate authority to execute. Fortunately, the multicellular design that demands such integration also provides a special class of cells to accomplish it. These cells— neurons—now do what Paramecium could not: provide multiple fast lines for communication. In short, for a multicellular organism a brain becomes necessary, possible, and profitable.

In some sense, the ability to adapt to your environment is intelligence.

The principles of neural design

I really love this. What are the simple principles from which brains emerge?

For designs to have persisted across this immensity of time and spatial scale implies that they are neither arbitrary nor accidental. Rather, they must have emerged as responses to some broad constraint. That is what elevates the shared responses to the status of principles.

Ok, so they define a ‘principle’ as a conserved feature across time and space, (and possibly evolutionary time and space).

send only what is needed; send at the lowest acceptable rate; minimize wire, that is, length and diameter of all neural processes

Ok, so what should these ‘principles’ explain? What can we do with principles and how can we use them to understand the brain?

The conservation laws should explain why neurons branch, what limits their individual volumes, their aggregate volumes in local circuits, and their fractional volumes (wire vs. synapses). The same laws should explain specialized substructures of local circuits, such as cortical layers, columns, stripes, pinwheels, and barrels—plus the structure of tracts. At a still larger scale, the laws should explain distinctive patterns of cortical folding, the relationships between cortical areas, and hemispheric specialization.

Here is a more verbose version of the principles.

Compute with chemistry

Bits per joule approaches lower bound set by thermodynamic limit.

Bits per liter reaches lower bound set by protein structure.

• Signals are fast at short distances.

• Computation is direct.

Compute directly with analogue primitives

• Analogue completes a basic operation in fewer steps than digital.

• Analogue is well suited to chemical and electrical computing.

Combine analogue and pulsatile processing

• Analogue processes information at high rates.

• Analogue electrical signals are cheaper than pulses.

• But stochasticity at all stages (vesicle release, ligand binding, channel opening) accumulates noise.

• Therefore, compute locally in analogue; threshold to restore S/N, and send noise-resistant pulses.

Sparsify

• Signal with proteins in small clusters.

• Release vesicles in brief bursts.

• Fire spikes in brief bursts.

• Maximize information per array for least space and energy: optimize fraction of active neurons; optimize S/N vs redundancy.

Send only what is needed

• Reduce noise and redundancy.

• Sculpt message for downstream users.

• Reduce number of signals to save energy and space.

Send at the lowest acceptable rate

• Higher rates cost disproportionately more.

Minimize wire

• Space and energy decrease as length and (diameter) 2 .

• Small diameter allows few bits per second.

• Slowest signals can use zero wire (neuromodulators, hormones).

• Shorter wires reduce processing time.

To shorten wire:

• Organize neurons in maps;

• Within a map segregate computations in parallel circuits.

• Separate circuits in layers, columns, stripes, barrels.

• Arrange maps to interconnect with least wire

• Connect neurons by matching their axonal and dendritic meshworks.

• Reduce instruction set to send long distance.

Make neural components irreducibly small

• Smaller reaction vessel allows faster chemistry with fewer molecules.

• Lower membrane capacitance charges with smaller current.

• Nanoscale molecular components allow smaller axons and synapses.

Complicate

• Specialize molecules to...