Duyal Yolcu

Duyal Yolcu

Neuroscience tech tree

4 minute read

Published:

Neuroscience tech tree

Status: Chaotic draft, will hopefully become more coherent within the next 2 months

The future is already there, it’s just not widely distributed yet - attributed to William Gibson

This is part of the “Neuroscience under a quantitative progress studies perspective” project with Kevin Kermani Nejad. An inspiration is the Longevity tech tree in the last page of this document.

In the introduction, we attempted to classify what humanity would like to understand about nervous systems to “reverse-engineer” them. In this post, we attempt the same for possible scientific and engineering advances towards these goals. We try to validate this classification in section XY, where we consider how well past advances in the field would have fit into it.

Apart from the classification itself, another implied assumption for it to be useful is that - as in the quote above - progress will come through technologies that are at least conceivable today, and become practically feasible through a rising tide of engineering progress and available experimental data and experience - rather than “one weird trick”-style advances that were hindered by nothing but noone thinking of them before. We will validate this assumption in section XY as well.

These techniques are interrelated. A focus of the current version is on the microscale - i.e. manipulating and understanding cells and lower-level structure. My current guess is that great breakthroughs in understanding nervous systems will rather be achieved by understanding smaller systems very well (than by understanding larger systems a little bit better), though this is the sort of thing to gather more substantiated opinions and evidence about.

This documents lives off interactions with experts and will be updated accordingly.

1. How can experimental methods improve?

An experiment is an interaction with a physical system - by which we include a computer or other form of simulator. Assume that, in all cases, the system is supposed to represent something happening in an actual brain. There is variety in which processes we are trying to represent; for any choice, we may improve (by quality/accuracy and/or cost)

  1. transfer of data into the system,
  2. the system itself
  3. transfer of data out of the system.

2. Types of systems

This is similar to section 2 in the introduction post.

  1. proteins/chemicals
  2. (interaction networks in) cells
  3. neurons and synapses
  4. neural networks
  5. nervous systems.

1. Proteins/chemicals

We are interested in evaluation - and reversing - the function that maps from a DNA sequence (+whatever further information is necessary) to the protein structure, and the reaction rates - and the reverse function (i.e. given a desired protein function, we’d like to find a DNA sequence).

We’d also like to know where proteins are located in a cell.

“Writing” is comparable to putting DNA into cells and making them express proteins. This is covered by DNA writing progress/the next section. A caveat is that

Understanding a given protein structure/”reading”: The Protein Data Bank is a databank of known protein structures; this section of the Wikipedia page on the Protein Data Bank) contains a statistic with a list of experimental methods for structure determination.

Reverse problem/”writing”: Not yet discussed

2. Reaction network inside a cell

  • understanding which genes are expressed (by GFP etc.??)
  • DNA ticker tapes, ticker tapes getting out/inside of cells…?
    • writing
      • gene editing:
        • targeted
        • random: observing phenotype changes of randomly induced mutations in population of cells
      • optogenetics/manipulating behaviour/gene by input
      • RF transmitter in cell?

3. Whole neurons

  • voltage sensitive dyes
  • patch clamps
  • synthetic extracellular environment neurochemicals to understand cell’s behavior
  • microscopes+cameras
  • calcium imaging (also for “nervous systems”)

4. Nervous systems

  • neuroimaging methods: everything on this list. See here for a review (from 2013) of physical limits of the main methods of the time; according to Milan Cvitkovic, ultrasound methods became more important in the meantime.

  • making nervous systems easier to interact with
    • direct genome manipulation
    • breeding
    • metabolism (lower -> waste heat less problematic)
    • physical space for recording devices
    • make axons grow towards electrodes themselves
  • data from human/animal populations
    • GWAS
    • ticker tapes, ticker tapes in bloodstream…

  • better understanding which information can be abstracted away

  • simulation/data processing
  • Advances in computing
    • Moore’s law/semiconductor advances
    • neuromorphic computing
    • analog computing
    • biocomputing/laboratories in cells
    • quantum computing
    • automated/robot scientists
    • ML for data-processing/statistical inference more generally
    • training data generation (as in OpenAI Gym)
  • more individualizable semiconductors (so that one-off/small-scale systems improve relative to industrial-scale systems)

3. Validation and caveats

todo…

You May Also Enjoy

Types of stochasticity and errors brains need to deal with/ways in which proteins in water are suboptimal for computation

6 minute read

Published:

Together, these may add orders of magnitude to the complexity and resources a brain has to use to accomplish tasks.

  • mutations in evolutionary history - preferable to have a brain structure surviving some mutations
  • diffusion times of neurotransmitters
  • stochasticity of reactions facilitated by proteins
  • protein folding: to implement the equivalent of a logic gate, one needs to invent some protein that folds in just the right way to facilitate a computation. That doesn’t sound easy at all, may introduce overhead
  • leakage (of current through axon/neuron walls)
  • stochasticity of neuronal growth and arborization?
  • Energy:
    • using too much energy is a much bigger concern for animals than for computers, so it is plausible that there are tradeoffs towards using less energy vs. better performance
    • relatedly: Mitochondrial volume limits energy inflow into brain; power density of biomatter is much lower than achievable in e.g. microchips
    • warm-blooded brain temperature can’t rise more than a few K before damage, limiting sustained energy expenditure. Can’t cool with liquid nitrogen either…
  • Evolution: There must have been a continuous evolutionary path from a bacterium to the human, or any other, brain. This probably doesn’t mix well with many “brittle” algorithms typical in CS, like cryptography, compression, maybe error correction…
  • because proteins and neurons are so incredibly slow compared to transistors, algorithms that can’t be parallelized well may not even be worth it
  • Indeed, there is little evidence for the sort of algorithms computer scientists devise (as in cryptography, compression, …) in biology. The most involved algorithm I heard of is the use of “modified Bloom filters” in flies remembering odors.
  • Parasites

Is your brain more complicated than your fridge?

6 minute read

Published:

The human brain is often described as “the most complex object in the universe” - justified by the number of synapses it contains or similar. But a glass of water contains lots of molecules and degrees of freedom as well.1 While the amount of relevant computation a brain performs during its lifetime may be high,2 the amount of information needed to describe it is upper-bounded by how much relevant data

  • was contained in the fertilized egg it developed from, and
  • flowed into it after conception.

  1. This problem is not just hypothetical: As of 2016, surrounding water molecules were a major expense in biochemical simulations according to what I heard. I don’t know the current status (QUESTION). ↩

  2. See Joseph Carlsmith’s report for a thorough attempt at estimation. The number of operations needed may be much higher or much lower than the number of spikes occuring in a natural brain. Furthermore, as they become more complex, artificial and natural computation systems tend to become bottle-necked by communication rather than computation - so the “number of operations required” may turn out to be irrelevant. TODO either elaborate, or remove this? ↩

Reachable states in quantum phase estimation

5 minute read

Published:

A famous, basic algorithm in quantum computing is the quantum phase estimation algorithm. We can see the algorithm as a quantum query algorithm 1 for oracles $O_d=\ket{\mathrm{idle}}\bra{\mathrm{idle}} + d\ket{\mathrm{v}}\bra{\mathrm{v}}$, where $d\in\mathbb{C}, \lVert d \rVert = 1$ is the eigenvalue to be estimated - for discretization purposes, we choose $|D|\in\mathbb{N}^{+}$ and assume $d$ is a $|D|$th root of unity, $d\in D$ with $D:=\left\{d\mid d\in\mathbb{C}, d^{|D|}=1\right\}$.

  1. https://github.com/qudent/RhoPaths ↩