< Previous | Home | Up | Table of Contents | Next >

Soft Uploading:

Outline of Soft Uploading v2.2

Soft uploading is a non-destructive variant of "standard" uploading and somewhat akin to the "Moravec" procedure (cf Joe Strout's web site). Claim: 1) Soft uploading is by most common-sensical standards a true implementation of physical immortality; and 2) It's probably the easiest approach (among other non-destructive methods) to implement.

Note: Many details need to be worked out. The outline will be subject to a lot of revisions that depend on, among other things, what kind of technologies will be available in the future, and the sequence of events.

The 2.0 version is much more realistic in terms of the design of DNI and the neuromatrix. I think these two can pass as 'technically feasible'. BSS remain relatively unexplored.

4-Part System

The Isolated Brain Support System (BSS) provides nutrients and oxygen to the brain via perfusion with artificial blood. The set-up may also include the spinal cord, a synthetic body, or an avatar (ie a remote body separate from the biological / uploaded brain).

The Direct Neural Interface (DNI) needs to penetrate the neuropil and contact cell bodies of neurons, allowing recording of membrane voltages as well as chemical information such as neuronal cell types or types of neurotransmitters. The DNI should also interact with the brain by sending signals to it (bidirectional transmission).

Non-invasive 3-D microscopy is used to locate neurons in the brain so that the DNI can establish connections with a high percentage of them. Alternatively, connectivity information can be extracted by molecular nanotechnology (outlined in the DNI page).

The Neuromatrix is just a name for a high-performance computer. Its main purpose (for now) is to gradually absorb information from the brain. Over an extended period of time, the brain deteriorates due to inevitable, biological ageing. The "center of consciousness" gradually shifts from the brain to the neuromatrix. Once the transfer of information is complete (meaning a significant amount of information in the brain has been transferred) then the "person" will live on the new substrate.


1) BSS: The idea of an isolated brain support system is not new. Currently there are a few medical companies working on the problem of short-term brain perfusion (eg for a few hours). Artificial blood is also an area of active research and some of it has entered clinical use. The prolonged sustaining of the brain would be more involved. One major problem is hormones in the circulation: which hormones must be provided by the blood substitute and which ones may be omitted. There are ~200 known hormones. More details on BSS.

2) DNI: Each probe is an insulated microwire (of ~0.1µm diameter) attached to a 'head-complex' that makes use of a magnetic field to penetrate the neuropil (aided by microscopy to locate neurons). When a target neuron is contacted, an electrode will be inserted into the neuron.

The complex also injects into each neuron molecular tags that allow the DNI to determine which neurons are contacted by that neuron's axon(s).

For more details see the DNI page.

3) Non-invasive Microscopy: The resolution of this microscopy need not be very high.

4) Neuromatrix: This is the hardest part. Basically we're faced with ~10-100 billion parallel channels of spike train sequences from which neural information (synaptic weights) has to be extracted. Massive parallel distributed processing is required. The neuromatrix operates on the idea of asymptotic convergence: synaptic weights are extracted by an algorithm similar to the Perceptron learning algorithm, which has a low-order polynomial time complexity in n = number of synapses. Mathematically each Perceptron is a hyperplane dividing a hypercube whose coordinates are synaptic weights. The learning algorithm progressively adjusts this hyperplane according to the sign of the error of the output. In fact, there exists a plethora of methods (from adaptive modeling, statistical learning, etc) that can be applied to this end (I have yet to sort this out, in the roadmap of machine learning).

Another problem is to decide on the parametric model of biological neurons. We may apply the statistical technique of dimensionality reduction to arrive at a model with the minimum number of independent parameters. Recent evidence in neurobiology suggests that such a model is likely to be nonlinear (unlike the Perceptron). Learning algorithms for nonlinear models usually takes the form of gradient-descend.

Then, armed with connectivity information (extracted by the DNI), the neuromatrix can be expected to extract a siginificant fraction of the brain's content in one or a few decades (optimistic guess), using the brain's own activity.

The above analysis rests on one assumption: Neurons communicate with each other exclusively through spike trains, with the exception of volume transmission (which includes gaseous diffusion). By Dale's principle we may assume that one neuronal sub-type release one blend of neurotransmitters / neurohormones. Thus the totality of neuronal information can be captured by spike trains plus the blend of neurochemicals of the cell-type in question.

For more details see the Neuromatrix page.


< Previous | Home | Up | Table of Contents | Next >

Feb/2003, Aug/2003, Dec/2003, Feb/2004 Yan King Yin