Whole-brain emulation: a roadmap, not a technology
Whole-brain emulation is a proposal to computationally reproduce a brain's functional dynamics from its measured structure. The story bounds the costly error: It is a hypothetical agenda with uncertainty about required detail, measurement, compute, validation and identity, not an available technology. It teaches how to turn an emulation promise into a chain of independent requirements and tests.
Whole-brain emulation is the proposal to reproduce, on a computer, the functional dynamics of a specific brain from its measured structure. It is not a project under way, nor a product: the document that defines it is a roadmap published in 2008 by Anders Sandberg and Nick Bostrom as technical report #2008-3 of the Future of Humanity Institute at Oxford University, drawn from a workshop held there on 26 and 27 May 2007 with specialists in computational neuroscience, brain scanning, computing, nanotechnology and neurobiology.
Knowing that changes how a headline reads. What exists is not a half-built machine but a map of what would have to be achieved — and one written under a rule its authors impose on themselves: list the uncertainties and propose experiments that would reduce them. They call it «falsifiable design», borrowing the idea from Nick Szabo: a theoretical technology deserves credit only if it pairs its untested claims with a plan for testing them. The report presents itself as a first draft, meant to be corrected as better information arrives.
The chain you have to walk end to end
The roadmap breaks the problem into three capabilities: physically scanning a brain to acquire the information, interpreting that data into a software model, and having the computing power to run it. Each drags its own requirements along. Scanning demands preparing the tissue —separating, fixing, staining it— and sectioning it into manageable pieces such that the same cells and dendrites can be matched on both sides of the cut; these are routine techniques in neuroscience, but here the volumes are far larger and the tolerance for damage far smaller.
Interpretation is an image-processing problem: correcting distortions, removing noise, filling in lost data, tracing cells, identifying synapses and cell types, and estimating the parameters the simulation will need. The output is not a picture but an inventory database of the scanned brain. Only then does computational neuroscience come in, which must supply good enough models of every relevant subsystem, and only at the end does the machine that runs it appear. The authors add three supports they consider indispensable: validation methods built into every step, low-level understanding to construct the models, and industrial-scale automation to handle the volume of data.
The level of detail decides everything else
The central question is not «can it be done?» but «at what resolution?». The report lays out eleven levels of emulation, from the abstract computational module to the quantum level, passing through region connectivity, analogue population models, spiking neural networks, electrophysiology, metabolome, proteome and the stochastic behaviour of individual molecules. Each level changes the nature of the problem: low levels demand enormous amounts of simple information; high levels little information but very complex.
At the workshop an informal poll of the attendees put the consensus at levels 4 to 6 —from the spiking network to the metabolome— with two participants more optimistic and two arguing that elements of levels 8 or 9 might be needed. To emulate at that level, the consensus was that a scanning resolution of 5×5×50 nanometres would be required. There is a physical reason: synaptic spine necks and the thinnest axons can be on the order of 50 nanometres or less.
The document also separates success criteria, and that table is what helps most in reading headlines. A parts list and a complete scan are valuable milestones, but they are not an emulation. Above them sit the brain database, functional emulation —producing at least a substantial range of species-typical basic emergent activity, such as a slow-wave sleep state or an awake one—, species-generic emulation, and only then the emulation of an individual brain, which would have to retain most of that particular brain's memories and skills. The three highest criteria —filling a social role, producing subjective mental states, and counting as a continuation of the person— are expressly left outside the analysis as poorly understood and hard to operationalise.
The assumption holding it up, and what would topple it
The whole agenda rests on one specific hypothesis: scale separation. That somewhere between the atomic and the macroscopic there is a level above which the slower, larger dynamics stop being sensitive to the faster, smaller ones. If that cut-off exists, measuring and simulating above it is enough. If it does not, whole-brain emulation is severely limited or infeasible, because no simulation of one particular scale will reproduce the causal structure of the whole.
From this comes a qualification the report itself stresses and that is worth keeping: a successful emulation need not predict every detail of the original system's behaviour; it need only replicate the computationally relevant functionality. Establishing where that boundary lies is, in the authors' words, a basic-science question to be settled with small emulations compared against the real system — not with declarations.
The numbers, and what they are made of
The report sizes the problem level by level. Emulating a spiking neural network would involve on the order of 10^11 neurons —a hundred billion— and 10^15 connections: some 8,000 terabytes of memory and a demand of 10^18 floating-point operations per second. Moving up to the electrophysiological level takes memory to about 10,000 terabytes and computation to 10^22 operations per second.
Beside each figure sits a year, and that is where the small print almost never quoted lives. Those years —2019 for level 4 on the optimistic supercomputer estimate, 2042 on the commodity-hardware one; 2033 and 2068 for the electrophysiological level— are not predictions about neuroscience: they are arithmetic on hardware prices, computed assuming Moore's law continues unchanged, and with the explicit caveat that one more order of magnitude of complexity adds only about five years to the tally. The authors give the slopes they use: memory per dollar rises one order of magnitude every 4.8 years and processing power per dollar one every 3.7 to 6.4 years depending on the series chosen. And they add the inverse exercise: a billion-dollar effort would bring the dates forward by eleven to nineteen years.
The conclusion they sign is conditional and should be read as such: if electrophysiological models were enough, full human brain emulation should be possible before mid-century, with simple mammals one to two decades earlier. Scaling to other species is a matter of synapses: the macaque brain has 14% of human synapses, the cat 3%, the rat 0.26% and the mouse 0.1%.
Why a small circuit does not prove the chain scales
The authors see no obstacles to attempting the emulation of an invertebrate organism today, and suggest starting with small, well-characterised systems. But the example they themselves pick shows the limit of the argument: the nervous system of Caenorhabditis elegans, with its 302 neurons, has been completely mapped since the 1980s and we still lack detailed electrophysiology for those neurons, most likely because investigating such small cells is so difficult. Having the map is not having the model.
A small circuit is useful for testing reconstruction methods and for discovering the level at which one must simulate; it does not show that the whole chain works on a human brain. And that distinction —between validating a technique and validating a scale— is exactly what gets lost when a laboratory demonstration is reported as a step towards full emulation.
Validation, where the report is hardest on itself
The document argues that neuroscience still works in a «debugging» paradigm: things get tested by replication or when something unexpected happens, without a systematic method. For a project of many chained steps that is not enough, because bad data at the scanning stage contaminates everything downstream. They propose two concrete tools. One, a gold-standard model at the highest feasible resolution against which to measure how far one can deviate before noticeable effects appear. The other, manufactured datasets whose ground truth is known in advance —generated, for instance, by a system that models neurite outgrowth and produces virtual slices— to check whether reconstruction methods recover the network that was actually there.
Emulation, simulation and other things that are not the same
The report fixes the vocabulary precisely. A simulation mimics outward results; an emulation mimics the internal causal dynamics, and so the term is reserved for a one-to-one model in which all relevant properties exist, whereas in a simulation only some do. Comparing whole-brain emulation with neuroscientific simulation, with neuromorphic hardware or with cognitive models means comparing objects with different success criteria: each is judged on its own terms, and the comparison only holds if it preserves the task, the budget and the cost of error.
The capability the reader takes away
What is usable in this document is not its calendar but its method: it turns a promise into a chain of requirements and independent tests. Faced with any brain-emulation announcement —today or ten years from now— five questions bring the conversation back to checkable ground. Which success criterion is being claimed: a scan, a database, species-generic activity, or a recognisable individual brain? At what level of detail was it simulated, and at what scanning resolution? Which scale-separation assumption is being taken for granted, and what experiment would test it? Where does the date come from: if it follows from an assumed hardware curve, it is arithmetic, not a forecast about brains. And how was it validated: against which reference, and with what known ground truth?
That list outlives the document that inspired it. The roadmap is nearly two decades old and its dates must be read with its own authors' caveat; the practice of demanding criterion, level, assumption, origin of the figure and validation does not expire, because it depends on no particular technology. It is also why this report is still cited: not for getting the calendar right, but for putting in writing what it did not know.
This article was produced with artificial intelligence under human editorial oversight.
