Are physical 3D artificial brains the next step in AI?

Data and algorithm based architectures have led to massive leaps in AI. But what if the next step in AI is a revolution in the substrates these run on?

Image: ChatGPT/Dall-E and Photoshop

Nine years ago I published a highly speculative commentary in the journal Nature Nanotechnology that considered the challenges and opportunities of 3D printing an “artificial mind”.

The article focused on the confluence of 3D printing, neuromorphic computing, and nanotechnology, as it explored the possibility of making massive strides in AI-optimized compute capabilities using dense three-dimensional structures.

At the time it was so edgy that my editor at Nature Nanotechnology was unsure whether the commentary was appropriate. However, nearly a decade on, it’s looking increasingly prescient! That said, we’re still only scratching the surfaces of compute substrates that could elevate AI to the next level and beyond.

I’ve included the text of the 2014 Nature Nanotechnology commentary below. Even though it’s a little outdated, it remains relevant to the opportunities and challenges that 3D AI substrates present. Before I get there though, it’s worth adding some additional context.

When I wrote the article, there was growing interest in neuromorphic computing — using hardware to mimic human-like neurons and synapses. But much of this was still at the research stage, and there was very little focus on stepping up from two dimensional to three dimensional chip architectures.

Now, a number of manufacturers are working on neuromorphic hardware. IBM, for instance, recently released its NorthPole brain-inspired chip. And a couple of years ago Intel launched their Liohi 2 neuromorphic chip.

There’s also growing interest in increasingly novel neuromorphic architectures. Just recently research was published on using arrays of silver nanowires as a neuromorphic compute substrate. And the company Rain Neuromorphics is purportedly building chips with analog components designed to emulate biological neurons.

These and other developments underline growing interest in accelerating AI through novel physical compute substrates. But most developments still rely on dimensionally constrained architectures. There are some projects focused on developing 3D chips — including the European NimbleAI project. But these are few and far between. And using 3D printing to create three dimensional non-biological neuromorphic compute architectures is, as far as I can tell, still in the realms of speculation.

Yet 3D neuromorphic compute architectures would be a game changer in terms of compute power, speed, compactness, and efficiency, if they could be manufactured.

This was the crux of the 2014 commentary — along with the challenges and opportunities associated with achieving functional architectures.

In many ways, the article is more relevant now than it was nine years ago. It misses some things — the need for a secondary network to systematically adjust weights between neurons for instance. But it still maps out possibilities and opportunities that we haven’t reached yet, but we’re getting closer to achieving in the race to develop ever more powerful AI.

Here’s the article in full (also available here):

---

Could we 3D print an artificial mind?

Nature Nanotechnology 9, 955–956 (2014) DOI: 10.1038/nnano.2014.294

3D printing is allowing more complex three-dimensional structures to be manufactured than ever before. Could the convergence between this technology and nanotechnology eventually usher in a new era of artificial intelligence, asks Andrew D. Maynard.

In 2013, the Obama Administration launched the Brain Research through Advancing Innovative Neurotechnology (or BRAIN) Initiative in the United States. In the same year, the European Union announced the similarly focused Human Brain Project. Together, these initiatives map out a far-reaching vision for the future of neurotechnology; one that will undoubtedly lead to major breakthroughs in understanding how the human brain functions.

Both the BRAIN Initiative and the Human Brain Project are focused on developing the science and technology needed to prevent and treat brain disorders. Both also aim to stimulate advances in brain-inspired technologies. Yet without parallel advances in the ways in which processors are manufactured, progress here could be slow. Our brains are analogue devices that are intimately dependent on their physical structure. Their analogue behaviour and their three-dimensional complexity can be simulated to a degree with conventional digital computing technologies, but only up to a point. Breakthroughs in neuromorphic processors are beginning to overcome some of these limitations by combining very large scale integration (VLSI) with brain-like architectures. And the field is beginning to show how conventional fabrication and unconventional configurations can emulate very simple brain-like systems. Yet such architectures remain limited by the degree to which massive integration is possible using conventional technologies. Ultimately, artificial constructs with advanced brain-like behaviour will require a transition to highly complex three-dimensional structures that more completely mimic that of the human brain.

This is a monumental challenge. The human brain consists of tens of billions of neurons and hundreds of trillions of synaptic connections, all spatially intertwined down to the nanometre scale. Such complexity remains well beyond the reach of even the most sophisticated conventional fabrication facilities. In contrast, spatial complexity in additive manufacturing processes such as 3D printing is relatively straightforward to achieve. The spatial resolution of most current 3D printing systems is still a long way from the nanoscale. But as the resolution and sophistication of devices continues to improve, it is not beyond the bounds of plausibility that the technology's convergence with nanotechnology could pave the way for future neuromorphic devices that blur the boundaries between programmed functionality and intelligence.

If construction of an advanced artificial brain were possible — not as a simulacrum of a biological brain, but as a bioinspired processor — it would open radical avenues of technological innovation. Three-dimensional analogue computing substrates could respond to external stimuli more rapidly, more 'organically', and more affordably than current systems. They may well be a vital step towards high-functioning artificial intelligence. Yet while it's easy to speculate on the opportunities resulting from the convergence of 3D printing and nanotechnology, a key first step is exploring whether there is a plausible likelihood of such a convergence occurring in the first place.

Although much about the human brain's functionality remains to be discovered, emerging understanding could, in principle, guide the development of simple bioinspired three-dimensional analogue computational substrates. Building on current advances in neuromorphic technologies, first-generation substrates could conceivably consist of arrays of processing devices that are capable of simultaneously receiving many signals; integrating and processing these signals; retaining a memory of past events; modifying how future inputs are processes based on past history; and outputting signals to other similar devices. In other words, an array inspired by, but not necessarily emulating precisely, a neuron–axon–synapse network. Novel forms of artificial neurons such as memristor-based 'neuristors' (as one example) may also soon satisfy the nodal requirements of such a network. These would be ineffective however without parallel systems to provide power to the devices, and remove the excess heat generated by them. Integrating a three-dimensional network of power supply circuitry and heat removal channels into a massively interconnected three-dimensional array of signal processors is a highly complex challenge, and a limiting factor in the application of conventional manufacturing technologies. 3D printing on the other hand provides an intriguing and novel route to overcoming these limitations.

With the appropriate 3D printing technology, an array of a thousand, or even a million processing units, should be no more difficult to manufacture than one consisting of only a handful of units (given a printer that's large enough). In principle, there is minimal additional overhead cost in increasing complexity up to the spatial resolution limit of a 3D printer. This is a game-changer for constructing neuromorphic processing substrates. By printing such substrates layer by layer, there is every reason to suppose that, one day, it will be possible to construct a massively interconnected array of processing devices, complete with integrated power supply and heat removal networks.

At present, this remains a future aspiration. The technology does not currently exist to 3D-print nanoscale components similar to 'neuristors' for instance. Yet it's plausible that with evolutionary advances in multi-material 3D printing at nanoscale resolution, it will be possible to print the necessary components within the next 10 to 20 years. Once this hurdle has been overcome, connecting these processing 'nodes' and providing power to each unit should be relatively straightforward.

Removing excess heat from the resulting dense three-dimensional substrate presents a larger technological challenge. In the human brain this is achieved predominantly through blood perfusion. A similar microfluidics approach should be amenable to 3D printing, as the necessary spatially arranged flow channels could be laid down layer by layer with relative ease. Alternatively, a secondary network of nanoscale heat pipes printed directly into the three-dimensional substrate may be a viable option. In either case, although challenging, heat extraction from a 3D-printed processing substrate should be possible as nanotechnology and additive manufacturing continue to converge.

Of course, this is a naive thought experiment, and one that glosses over many substantial technological challenges. It does serve to illustrate however that, although still speculative, the spatial complexity achievable through 3D printing could plausibly lead to the formation of transformative neuromorphic processing architectures.

If the convergence of 3D printing and nanotechnology does enable the formation of highly novel three-dimensional computational substrates, they will conceivably form the basis for developing artificial brains that are inspired by our own, but do not necessarily emulate them. This could be a pivotal point in the development of advanced artificial intelligence, and a significant step towards realizing the as-yet science fiction imaginings of artificial minds. It's a future scenario that is both intriguing and challenging. The idea of 'mind' implies a degree of awareness of self and environment — an emergent property that some would argue is a result of the highly complex functioning of biological brains. This is a far remove from simply manufacturing a three-dimensional computational substrate that can be trained or programmed to analyse and respond to stimuli in specific ways. It's easy to assume of course that a 3D-printed 'brain' is no more likely to show mind-like properties than a powerful computer. Except that, the only experiences we have so far with massively interconnected three-dimensional analogue processing structures is within living organisms. And we know from experience that the complexities and emergent properties of these biological systems differ markedly from anything we've so far been able to achieve through two-dimensional-constrained processors and software emulation.

Which leads us to a question that is, if anything, more difficult to address than the aforementioned technical hurdles: if our technological capabilities are beginning to shift from the fanciful to the plausible in constructing an artificial mind that has some degree of awareness, how do we begin to think about responsibility in the face of such audacity?

Maynard, A. Could we 3D print an artificial mind?. Nature Nanotech 9, 955–956 (2014). https://doi.org/10.1038/nnano.2014.294