Neuroscience and the Hyperreal
Part 1 of a series on models in neuroscience
“… the logical evolution of a science is to distance itself increasingly from its object, until it dispenses with it entirely: its autonomy is only rendered even more fantastic – it attains its pure form.”
Jean Baudrillard, Simulacra and Simulation
When I think about the state of neuroscience these days, I can't shake the feeling that something is off.
The brain is becoming hard to discern in a lot of neuroscientific research. For a while I couldn't articulate how or why this might be happening. But then a single word popped into my head: hyperreality.
The French post-structuralist thinker Jean Baudrillard coined the term "hyperreal" to capture what he saw as a transformation underway in contemporary culture. In his 1981 book Simulacra and Simulation, Baudrillard defined the hyperreal as "the generation by models of a real without origin or reality".1
I would like to explore the idea that neuroscience is drifting towards a hyperreal epoch, when computational models may cease to be guides to reality. Instead, they will come to obscure and eventually "replace" reality.
Fleshing out this premonition requires explaining a few different moving parts, which I will do over the course of the next few posts.
In this introductory post, I'll sketch the broad contours of "neural hyperreality". In future installments, I'll pencil in some finer detail. I'll end the series with some suggestions for how the slide towards hyperreality can be countered by reframing the role of models in orienting scientific attention.
The rise of inexplicable models in neuroscience
To understand the road to the hyperreal we must trace the rising profile of models, which parallels the growth in the sizes of datasets available for modeling. These trajectories also mark the progressive blurring of useful distinctions between different types of model.
Each type of model in science has a different purpose and structure. In the next post I will explore a few different ways of contrasting model types, each of which draws out a different facet of the modeling process. Here I will introduce one of them: the contrast between statistical models and mechanistic models. Statistical models extract patterns from data, enabling the data to be summarized, classified, or fit using mathematical trend-lines. Mechanistic models simulate "mechanisms", which may be derived from well-established physics and chemistry theory, from coarse summaries of observations, or from speculative guesses about what is "actually" happening. Mechanistic models that don't have an obvious link to the physical sciences are often called "phenomenological models".
Ideally, statistical models tell you whether the data might be evidence of a repeatable phenomenon in the first place, while mechanistic models tell you how a phenomenon might arise from the interplay of real or hypothetical “forces”.
Mechanistic models have not been taken very seriously by experimentalists, who dominate the field of neuroscience. Experimental neuroscientists tend to treat mechanistic models as suspiciously speculative, if they are even aware of them. There is one very notable exception: the Hodgkin-Huxley model of the action potential, which is widely recognized as one of the most accurate and mechanistically insightful models in all of biology2. But for any scale larger than that of a single neuron, experimental neuroscientists have tended to rely exclusively on statistical models.
By the 2010s, datasets had started to grow rapidly in size and complexity, thanks to a crop of new experimental methods including fMRI and optogenetics. Many experimentalists found themselves confronting the limits of simple statistical models. More complex statistical models became feasible thanks to advances in computing power, but using these models correctly required a deep understanding of statistics.
To anyone who did have this level of understanding, it could often seem as if experimentalists were using statistical models as black boxes — data goes in, something mysterious happens, and publishable results come out. To add to the murk, every lab was using its own idiosyncratic mix of statistical models. Sometimes a research team would hop from model to model until they land on something "statistically significant"3.
And then deep learning burst on the scene, accelerating the proliferation of models.
Deep learning is a machine learning (ML) technique for training artificial neural networks (ANNs). ANNs have an unprecedented ability to perform pattern detection, classification, and function approximation. They seem to beat traditional statistical models at their own game4.
The success of ANNs doesn't come cheap, however. If you want the best possible results, you need to rely on the magic of scale. The performance of an ANN is often predicted by its size and by the size of the dataset used to train it. And the larger the model and dataset, the more computational power you need for training. This in turn makes state-of-the-art ANNs expensive and resource-intensive.
ANNs are also very hard to understand, even for the people who design them. There is no consensus about why they are so good at uncovering latent patterns, or why they fail in strange and unpredictable ways5. This has led to a whole cottage industry devoted to "explainability" and "interpretability" in AI and ML.
Inexplicable models are not really a new problem for experimental neuroscientists. The black box stance that many experimentalists were already adopting with respect to traditional statistical models is even more suited to ANNs, given that no one really knows how they work.
So to summarize, ANNs are efficacious, expensive, and enigmatic. And this combination of features generates an aura that makes the distinction between statistical and mechanistic modeling harder to recognize — particularly in a field where mechanistic modeling was never very popular in the first place.
Model muddle
Most ANNs used in ML are not really "neural" in a detailed mechanistic sense: they are based on networks of artificial neurons, and neither the neurons nor the networks they are embedded in bear a strong resemblance to their biological counterparts. And the techniques for training these models are even less biologically plausible than their architectures.
ANNs are nevertheless starting to be used in some circles as quasi-mechanistic models of the brain. The primary reason for this, apart from the fact that they are made up of artificial neurons, is that ANNs mimic human performance on a variety of tasks with far greater accuracy than previous generations of models that were more neurobiologically grounded. What this means is that for some neuroscientists, the primary metric to assess a model is statistical performance on a specific set of tasks, even if the model is otherwise implausible and/or impenetrable.
When ANN models match humans at the level of tasks, it is natural enough to ask if they match humans at the "neural" level too. For example, experimentalists have found correlations between activity patterns in image-processing ANNs and in the visual systems of the primate brain. There are various technical critiques that this approach has elicited6, but for our purposes the big problem is mechanistic-sounding extrapolation from such comparisons.
Researchers that are overly impressed with the quantitative metrics of statistical modeling — task accuracy and correlation with neural activity — may come to treat such metrics as a substitute for broader and more qualitative scientific goals, such as mechanistic understanding, explainability, and theoretical unification. These sorts of goals require a deep engagement with the experimental and theoretical literature.
Modern stats and AI tools nudge researchers in the opposite direction, making scholarly engagement seem old-fashioned and arduous. The tools are just so convenient to use! The popularity of ML methods in industry has led to a vast eco-system of software packages that are far more professional and user-friendly than the hacky tools that neuroscientists have come up with. These packages create a relatively frictionless experience that can induce researchers to search where there is available light, restricting their attention to the subset of experimental findings that are the most “machine readable”.
The hyperreal state will arrive if this process is taken to its logical conclusion: traditional conceptions of scientific understanding will then come to seem irrelevant in the face of the sheer power of the new models. Vast swathes of neuroscientific and psychological research will be ignored because they cannot easily be formatted as training data. In a hyperreal universe, many scientific questions will no longer be askable.
I think it is perfectly possible for neuroscience to avoid this fate. And this does not require rejecting any sort of model — so if you really love ANNs, rest assured that I am not insinuating that these models cannot be used productively in neuroscience. The risk of hyperreality does not come from any specific modeling approach, but from certain ways of using models to direct scientific attention. As I work my way to the end of this series of posts, I’ll offer some thoughts on how models can be used to enrich engagement with phenomena related to the brain and behavior, instead of displacing those phenomena in the minds of researchers.
Notes
The use of postmodern theory to make sense of AI/ML seems to be having a moment. A recent substack post by Henry Farrell riffs on a recent book linking “Theory” with large language models: Cultural theory was right about the death of the author. It was just a few decades early
I once joked that "LLMs are perfect Derridaeians - “il n'y pas de hors texte” is the most profound rule conditioning their existence.” Weatherby’s book provides evidence that this joke should be taken quite seriously indeed.
Interestingly, the Hodgkin-Huxley model is often regarded as the ground truth about how the flow of ions across a neural membrane gives rise to the action potential — despite the fact that even Alan Hodgkin and Andrew Huxley themselves knew that the model deviated from observations. This is a microcosm of the hyperreal replacement of reality by models.
A few years ago I gave a talk on the Hodgkin-Huxley model as part of a discussion series on dynamical systems in neuroscience. I charted the historical development of the model and explore the close integration between experimental findings and mathematics. Check it out here:
The Action Potential - from Galvani to Hodgkin & Huxley
It should be noted that statistical hygiene has improved quite a bit in recent years, but I have heard so many horror stories from friends in experimental labs that I can’t imagine the problem has been completely addressed.
The emphasis there is on seem. Plenty of people with a traditional statistical background express worries about the sheer number of parameters in ANNs: they seem to raise the risk of a phenomenon called overfitting, but for reasons that are still being actively studied, ANNs consistently defy the older statistical intuition.
Here’s an interesting example from a company called FAR.AI. It turns out that AI Go players with “superhuman” skills can have surprising failure modes. These vulnerabilities can be easily explained to human players, who can then use them to beat the AI players.
Noam Chomsky may have been describing ANN-based models when he said, “There is a notion of success… which I think is novel in the history of science. It interprets success as approximating unanalyzed data
Very interesting, and I look forward to future posts on this. What you're saying resonates with a similar observation about the fields of theoretical physics and cosmology. I'm seeing increased reliance on speculative models, and I worry that we may be drawing conclusions from inaccurate models almost out of desperation due to lack of real progress in the fields.
The notion of hyperreal seems to me to parallel my sense that culture and science are increasingly detached from physical reality. Both seem more and more wrapped up in sheer fantasy.