There's a story everybody in ML has heard. Deep nets shouldn't work. They have far more parameters than data points, so classical statistics says they should memorize the training set and generalize terribly. They can fit pure noise. They show weird curves where adding parameters first hurts then helps. They're a special, mysterious beast that breaks the old rules.
Andrew Gordon Wilson's position paper "Deep Learning is Not So Mysterious or Different" (arXiv:2503.02113) makes a deliberately contrarian argument: most of that story is wrong. The "anomalous" behaviors — benign overfitting, double descent, the success of overparametrization — are not unique to neural networks, and they're not mysterious. They follow from generalization principles that have been around for decades. You can characterize them rigorously with PAC-Bayes and countable hypothesis bounds.
And then, having deflated the hype, he's careful to say where deep learning is genuinely special. The result is one of the more clarifying reads in the whole pile — less a new method, more a new pair of glasses.
The three "anomalies" on trial
Benign overfitting
A model fits the training data perfectly — including the noise — yet still generalizes well to new data. Classical intuition says fitting noise must hurt. Deep nets seem to break this.
Wilson's point: benign overfitting shows up in other model classes too (linear models in the right regime, kernel methods, Gaussian processes). It's a property of flexible models with the right inductive bias, not a neural-network quirk. A model can fit the noise locally while still preferring a globally simple solution everywhere else.
Double descent
Plot test error against model size. Classically you expect a U-shape: too small underfits, too big overfits. Instead you often see error go down, then up near the point where the model can just barely fit the data (the "interpolation threshold"), then down again as the model gets even bigger. Two descents. Spooky.
Wilson: also not unique to deep learning — double descent appears in linear regression and random feature models. The bump is about how the model behaves right at the interpolation threshold, where it's forced to fit exactly with no slack; give it more capacity and it has room to find a smoother, simpler interpolant. Nothing neural-specific is required.
The success of overparametrization
More parameters than data, yet better generalization. The headline "paradox." Wilson's reframe is the heart of the paper.
The number of parameters is the wrong thing to count. What matters is the model's effective complexity — its soft preference for simple functions — not its raw parameter count. A huge model that strongly prefers simple solutions is, effectively, simple.
The unifying idea: soft inductive biases
Here's the key concept, and it's worth internalizing because it reframes everything above.
The old way to avoid overfitting: restrict the hypothesis space. Only allow simple models (low-degree polynomials, few parameters). A hard constraint — simple solutions are allowed, complex ones are forbidden.
The soft inductive bias way: allow a flexible hypothesis space — let the model be able to represent very complex functions — but bake in a soft preference for simpler solutions that fit the data. Don't forbid complexity; just lean toward simplicity when simplicity explains the data.
Fig 1 — The shift Wilson highlights: don't shrink the hypothesis space to forbid complexity (left); keep it flexible but softly prefer simple solutions consistent with the data (right).
This single principle dissolves the paradox. An overparametrized network can represent wild functions, but its training dynamics and architecture give it a soft pull toward simple ones. So it generalizes like a simple model while retaining the flexibility to fit when the data genuinely demands it. And — Wilson's main point — this principle isn't neural. You can encode soft inductive biases in many model classes. Hence: not so different.
optional read — why PAC-Bayes captures this
PAC-Bayes bounds the expected test risk of a posterior $Q$ over hypotheses by the empirical risk plus a complexity term driven by the KL divergence from a prior $P$:
$$\mathbb{E}_{h \sim Q}[R(h)] \;\le\; \mathbb{E}_{h \sim Q}[\hat{R}(h)] + \sqrt{\frac{\mathrm{KL}(Q \,\|\, P) + \ln\frac{1}{\delta}}{2n}}$$
The bound doesn't count parameters. It counts how far the learned solution sits from a simplicity-favoring prior. A flexible model that lands on a low-complexity solution (small KL to a simple prior) gets a tight bound regardless of how many parameters it nominally has. That's the rigorous version of "soft inductive bias," and it's exactly why countable hypothesis / PAC-Bayes frameworks — decades old — already explain the "anomalies."
Count effective complexity, not parameters
The crux deserves its own example, because "parameter count is the wrong measure" sounds glib until you see it. Consider two models with a billion parameters each. Model A is a billion-param network with no regularization, trained to wiggle through every noisy point. Model B is a billion-param network whose training dynamics and architecture pull it hard toward smooth, low-frequency functions. They have identical parameter counts. Yet B generalizes and A may not. The parameter count predicted nothing; the effective complexity — how much wiggle the model actually commits to — predicted everything.
This is why "more parameters than data" was never the paradox it was sold as. A flexible model that prefers simple functions is, in any sense that matters for generalization, a simple model that happens to be able to be complex if the data forces it. The capacity is latent, not spent.
The same lens dissolves "neural nets can fit random labels." Of course they can — they're flexible. The point is that on real (structured) data they don't choose to; the soft bias steers them to the simple explanation when one exists. Capacity-to-memorize and tendency-to-memorize are different things, and conflating them is the root of the "mystery."
The Occam connection
There's a clean way to see why a soft preference for simplicity generalizes: it's Occam's razor with a probability attached. Simpler hypotheses are a priori more likely (there are fewer of them, they need fewer bits to describe). A learner that places more prior mass on simple solutions, but still permits complex ones, will — when several hypotheses fit the data — land on the simplest, which is the one most likely to keep working on new data. PAC-Bayes makes this exact: the complexity penalty is literally how far you moved from a simplicity-favouring prior. No appeal to anything neural.
So where IS deep learning actually different?
This is what saves the paper from being mere contrarianism. Wilson is explicit that deep learning is distinctive — just not in the ways the folklore claims. The real differences:
- Representation learning. Deep nets learn their own features from raw data, building hierarchical representations. This is genuinely powerful and not something kernel methods or linear models do in the same way. The generalization story is shared; the feature-learning story is not.
- Mode connectivity. Different trained solutions (different loss minima) are often connected by low-loss paths in weight space. The loss landscape has a structure that's special to these models and still being mapped.
- Relative universality. The same architecture and recipe work across wildly different domains — vision, language, audio, protein folding. That breadth is unusual and is arguably the practical reason deep learning ate the field.
The clean takeaway: generalization in deep learning is not mysterious — old theory explains it. What's special about deep learning is representation learning, the geometry of its loss landscape, and its sheer universality.
Why this paper is worth reading even though it has no method
It's a position paper, not an experiment dump, and that's the point. The value is conceptual hygiene. So much commentary treats deep nets as a magical exception to statistics, which leads to bad intuitions and wasted effort chasing "new" theory for phenomena old theory already covers. Wilson's reframing — count effective complexity, not parameters; think soft preference, not hard constraint — is the kind of mental correction that makes you read the next twenty papers more clearly.
The fair pushback: "explicable in principle by PAC-Bayes" is not the same as "we have tight, predictive bounds for real networks." The frameworks are the right language; turning them into quantitative predictions for a specific 70B-parameter model is still open. Wilson is arguing about the nature of the phenomena, not handing you a calculator. But as a corrective to the "deep learning breaks all the rules" reflex, it lands.
arXiv:2503.02113 — Deep Learning is Not So Mysterious or Different, Andrew Gordon Wilson. Presented at the YC Paper Club.