The world is too loud. Read what matters.

Many Minds

Deep learning predicts neurons better than linear models, but nobody can explain why

The speaker argues that AI models create a sharp divide between prediction and understanding in neuroscience—the more accurate the model, the less scientists can explain why it works—exposing perhaps an insurmountable limit to scientific understanding itself.

NeurosciencePhilosophy of ScienceAI InterpretabilityDeep LearningComplex SystemsEpistemology
This is a rigorous philosophy of science lecture, not an interview. The argument is solid, with concrete examples—her own PhD research serves as a counterexample—but the language is academic. It suits patient readers concerned with AI interpretability.

The argument · tap a timestamp to hear it

7:20

Linear models dominated visual neuroscience until contradictory data emerged

Chirimuuta's advisor David Tolhurst had tested since the 1970s whether primary visual cortex neurons operate as linear filters: measure response to simple stripes, predict response to any image. By the 1990s, data contradicted this model—neurons were influenced in unexpected ways by their neighbors.

— Mazviita Chirimuuta
11:24

Deep convolutions predict far better than linear models, but as a black box

By 2018, when Chirimuuta revisited the problem after sabbatical, many labs had switched to deep convolutional networks (like those in face recognition) to predict visual neuron responses. These vastly outperformed improved linear models. But they rest on no functional hypothesis about what neurons actually do—pure black boxes yielding predictions without explanations.

— Mazviita Chirimuuta
12:25

Her advisor's email: understand it or predict it, not both

She shared the finding with her advisor David. His reply crystallized the dilemma: build a simple, mathematically transparent model you can understand but that only works narrowly, or build a complex, mathematically opaque model that predicts everything but yields no understanding. The conclusion: abandon understanding, and you're doing engineering, not science.

— Mazviita Chirimuuta (paraphrasing David Tolhurst)
15:27

AI exposes a hard limit to what science can understand

As a philosopher, Chirimuuta reinterprets this dilemma as evidence of science's hard limits. Not the ultimate mystery of consciousness, but this: even one visual cortex cell operates so intricately that scientists need mathematically complex AI models to predict it, models so complex their creators cannot understand them. The implication: some questions about nature may admit no humanly intelligible answer.

— Mazviita Chirimuuta
31:32

DiCarlo redefines understanding as the ability to control

MIT's DiCarlo lab used neural networks to design stimuli precisely controlling V4 neuron firing, countering claims that models are mere spectacle, not genuine brain understanding. The commentators noted this actually shifts understanding from ‘decomposing and explaining principles’ to pure pragmatism: if you control it, you understand it—as most people understand cars, predicting and controlling behavior without knowing how they work.

— Mazviita Chirimuuta
33:34

Poldrack: prediction and understanding are both compression at heart

Stanford psychologist Russell Poldrack rejects the dilemma: understanding and prediction are fundamentally identical—both compress messy data into compact form. Simple linear models (say, two-parameter fits) compress coarsely, discarding much information. Neural networks with many parameters compress high-dimensional data, retaining far more signal without overfitting. That's why they predict better.

— Mazviita Chirimuuta (paraphrasing Russell Poldrack)
39:42

Some scientists abandon reductionism but still pursue understanding

Rather than embrace pure engineering, some scientists preserve both understanding and intervention while abandoning mechanistic and reductionist models. Philip Ball and Nicole Rust argue past Alzheimer's explanations—a faulty protein sparking a cascade—were crude oversimplifications. Real pathology is a dense causal network across scales, riddled with feedback, causes and effects inseparable. Hence why clearing amyloid plaques shows almost no clinical effect on cognitive decline.

— Mazviita Chirimuuta
56:02

The mind-body problem may be an artifact of simplified science

Chirimuuta returns to Dubois-Reymond's premise: consciousness lies forever beyond science's grasp. She suggests the mind-body problem itself is an artifact of 17th-century science's insistence on modularity and clear boundaries. Once we grant that ‘pure physical’ is merely simplification and idealization, the mind-body split dissolves. But this dissolution yields only ‘useless knowledge’—understanding that generates no concrete interventions.

— Mazviita Chirimuuta

In their own words · checked verbatim

Either you build a simple model that makes mathematical sense and gives you understanding... or you can build an immensely more complex mathematically opaque model that works accurately across the board, but doesn't yield understanding of the responses.

Mazviita Chirimuuta12:25

This would mean that science with its central goal of understanding the natural world has a limit.

Mazviita Chirimuuta15:27

We understand a system when we can control it. Most of us understand our cars in this latter sense, predictability and controllability without really understanding them in the former reductive sense.

Mazviita Chirimuuta31:32

That there is not a trade-off between prediction and understanding... as I for one have argued because the concepts of prediction and understanding are linked both involve data compression.

Mazviita Chirimuuta33:34

The power of a technology is proportional to its inherent outofcrollness, its inherent ability to surprise and be generative.

Mazviita Chirimuuta38:41

We would have attained some useless knowledge.

Mazviita Chirimuuta56:02

Figures

Chris Anderson's 'end of theory' proposal200827:31
Scientific American cover questioning complexity science hype199554:01

Glossary

foundation models
Large general-purpose models trained on vast data, like large language models, transferable to many prediction tasks.
mechanistic interpretability
Research method that reverse-engineers AI function or behavior by analyzing internal activation patterns.
maker's knowledge
Understanding gained by building a system firsthand and knowing its construction principles; the source of traditional scientific understanding.
complex adaptive dynamical system
A system view emphasizing multi-scale feedback and irreducibility to simple causal chains, replacing mechanistic models.

How to listen

Who it's for

Researchers and engineers interested in philosophy of science, AI interpretability, and neuroscience modeling methodology.

Skip

The opening thanks for the Lakatos Award and speaker biography (0:13-7:20) can be skipped.