Cryonics and information theory

From: Eliezer S. Yudkowsky (sentience@pobox.com)
Date: Mon Jul 07 2003 - 02:05:20 MDT

  • Next message: Giu1i0 Pri5c0: "Re: Cryonics and information theory"

    Harvey Newstrom wrote:
    >
    > We have a lot better imaging technology now, and pictures of patients'
    > brains who have been frozen. The damage is a lot worse than we thought.
    > The cells shrank and pulled apart from each other with gaps between them.
    > Where the cells stayed put, they are disconnected from other cells. Where
    > the cells stayed connected, they are all pulled together leaving huge gaps.
    > The structure and positional relationship between the cells may or may not
    > be recoverable. Some people have likened this result to "hamburger", under
    > the analogy that resurrecting a brain in that condition would be like
    > resurrecting a cow from hamburger.
    >
    > I am still signed up for cryonics, but it seems to be to be a very low
    > probability chance of success, barely better than nothing. Many people
    > don't even bother to sign up for this reason. I certainly hope that better
    > methods can reduce this situation before my time comes.

    I've spoken with Peter Passaro about this, and apparently the main things
    are (a) get frozen as soon as possible after death (b) use the new
    oxygenated cryoprotectants (c) keep the brain from being starved of
    oxygen. So people are working on it, and if you get frozen with the
    latest techniques your chance of survival may still be pretty good.

    One problem I have, though, is that it still looks to me like it would be
    better to just chop off the head and drop it into a bucket of liquid
    nitrogen as fast as possible. *Large*-scale freezing damage is
    irrelevant; you can still connect the dots easily enough. What you want
    to ensure is that the information, the Shannon information, is still
    there. I would not be surprised to find even the earliest cryonics
    patients are resurrectable in toto; it is not necessary that the cells be
    reparable but that their physical state, when scanned down to the atomic
    level, contain enough information to extrapolate back the original brain
    and its relevant high-level information. The critical parameters here are
    a matter of information theory, not just medicine, and not at all obvious
    (i.e., how many initial states map to the same post-freezing state,
    whether critical information is in global patterns or local patterns,
    whether information makes a distinction in the final molecular state even
    if the apparent functional characteristics of the neuron have been destroyed).

    I worry that cryonics has been approached from the viewpoint of medicine
    rather than information theory. Here is a point where lack of optimism
    about post-Singularity capabilities may have killed people - cryonicists
    thinking "let's keep the neurons as undamaged as possible from the
    viewpoint of biological function" rather than "let's try and create a
    physical freezing process such that the configuration space of pre-frozen
    brains is mapped to the configuration space of molecularly analyzed frozen
    brains in a way that does not introduce information-loss on the level of
    relevant functional information". These are not at all the same thing;
    one is concerned not with how much "damage" the freezing process does,
    from the viewpoint of ice crystal formation and so on, but rather with the
    question of whether ice crystal formation of just dumping a head into
    liquid nitrogen is a physical process that maps many initial states into
    one final molecular-level state to a greater degree than the retraction of
    axons and dendrites that occurs if you leave the brain without oxygen.

    To give an example of how different the viewpoints are, slicing an area of
    neural tissue in half and translating one of the pieces by several
    millimeters is extremely destructive from a biological point of view, yet
    if the slice is a good one and the translation is consistent, almost no
    information has been lost - each point in the original configuration space
    maps to a unique point in the new configuration space. The question about
    ice crystal formation is not how much "damage" it does to the neurons, but
    whether as a physical process it tends to map distinct initial conditions
    onto distinct outcomes.

    If dendrites and axons retract into the cell body within half an hour
    after the neuron has been starved of oxygen (!!!), even so the essential
    information *may* have been preserved; the question is whether scanning
    the neuron on the *molecular level* would enable you to determine where
    the original dendrites and axons were, to a degree necessary to reproduce
    the functional information. In turn, you can only determine this by
    running several possible dendritic configurations forward in time under
    the retraction process, and seeing if several functionally different
    initial configurations map to exactly the same (molecularly the same)
    final retracted neuron. If the mapping is nonunique, however, you're
    probably toast, unless the gross position of neurons is a constraint
    sufficient to reconstruct the functionally relevant information of the
    original circuitry - if there is only one person you could have been such
    that your neurons would have occupied that gross position.

    What determines this? The degree to which precise details of the final
    post-freezing configuration constrain the original circuitry, and the
    degree to which the constraint is global in nature rather than local,
    relative to the functional space of brains. For example, suppose that in
    some area A1 we have a lossy mapping from a set of neural circuits N1 to
    the post-freezing brainstate F1. And suppose that the set of possible
    initial neural circuits N1 that map to F1 contain possibilities that are
    functionally different from each other. Are you dead meat? Perhaps and
    perhaps not. Suppose that there is Shannon information between the
    pre-freezing states of A1 and A2, such that if we know the pre-freezing
    state of area A1, it would constrain the permissible states or probability
    distribution of area A2. And suppose that, on a local level, there are
    many different circuits N2 that could have frozen to the final state F2,
    and some elements of N2 are functionally distinct from one another.
    However, there's only one possible element of N1 that is compatible with a
    possible element of N2, and only one possible element of N2 that is
    compatible with that particular element in N1. This is an idealized
    example; you can have probability distributions that constrain and narrow
    each other without this kind of definite certainty emerging from
    inspection of a mere two areas.

    The upshot is that if local areas of pre-frozen brains constrain one
    another (relative to the space of functionally different brains) in a way
    that survives mapping to frozen brains, such that local areas of frozen
    brains constrain information globally rather than locally, then even large
    local blurs may not destroy global preservation of information. If,
    however, local areas do not strongly constrain one another, then even a
    small local blur may permanently destroy your mind-state. All the locally
    uncertain probability distributions will modularly add up to an extremely
    uncertain global probability distribution, rather than many local
    uncertainties constraining each other to add up to global certainty. The
    greater the *locality* of the brain, in other words, the more easily it is
    destroyed by blurring.

    Blurring may be defined as mapping of functionally distinct local initial
    conditions to physically identical local final states; or more formally,
    mapping such that the densest volumes of the probability distribution for
    the final states tend to overlap one another even for functionally
    distinct initial conditions. Whether a given degree of local blurring
    kills you will depend on the degree to which those functionally distinct
    local initial conditions permitted by the final local physical state
    provide Shannon information about each other on a global scale, and
    whether that Shannon information can constrain the whole brain to a single
    functional state or whether it only narrows the blur without managing to
    eliminate it.

    It all boils down to the probability distributions for p(brain|mind) and
    p(frozen|brain), which together will determine p(mind|frozen). As usual,
    your life or death depends on - wait for it - Bayes' Theorem.

    So whether cryonics will work is a question that intersects biology,
    physics, and information theory, and the properties that determine whether
    you live or die are not at all obvious if you are thinking
    anthropomorphically about "preventing (biological) harm to neurons". What
    I worry about is not that cryonics has been misunderstood by the public,
    but that it has been misunderstood by cryonicists.

    Feel free to forward this message to Cryonet.

    -- 
    Eliezer S. Yudkowsky                          http://singinst.org/
    Research Fellow, Singularity Institute for Artificial Intelligence
    


    This archive was generated by hypermail 2.1.5 : Mon Jul 07 2003 - 02:15:46 MDT