Position: Deep Learning is Not So Mysterious or Different

Source

• Raw Markdown: paper_deep-learning-not-mysterious-different-2025.md
• PDF: paper_deep-learning-not-mysterious-different-2025.pdf
• Preprint: arXiv 2503.02113v2
• Proceedings: PMLR volume 267, pages 82326—82346
• Conference page: ICML 2025 Spotlight Poster
• OpenReview: ICML 2025 forum
• Official author X thread: Andrew Gordon Wilson, 2025-03-05
• Official author talks: Andrew Gordon Wilson talks page

The complete 12-post author thread and its 10 attached images were captured through the X API and are preserved in the raw paper directory as papers/deep-learning-not-mysterious-different-2025/x_thread_author_1897305433079161235.md, .json, a media directory, and x_thread_author_1897305433079161235_media_manifest.json. The manifest records post IDs, media keys, canonical URLs, local filenames, byte sizes, dimensions, and SHA-256 hashes. These auxiliary snapshots are narrative context and are not part of the public artifact surface.

No dedicated code repository was found during exact-title, arXiv-ID, proceedings, author-page, and X-thread checks. The paper is mainly a position and synthesis paper; several experiments and figures adapt or reproduce results from cited prior work.

Status And Credibility

Andrew Gordon Wilson first submitted the paper to arXiv on 2025-03-03 and revised it on 2025-07-10. It was accepted as an ICML 2025 Spotlight and published in the Proceedings of Machine Learning Research. ICML is a tier-1 machine-learning venue, and Wilson is an established NYU researcher in probabilistic machine learning, Gaussian processes, Bayesian deep learning, and generalization.

As of this ingest, the source is more than one year old. It remains credible as a peer-reviewed conceptual synthesis and generalization-theory position, not as a current state-of-the-art model or benchmark result. Its strongest role in this knowledge base is to calibrate how parameter count, interpolation, compression, and inductive bias are interpreted.

Core Claim

The paper argues that three phenomena often presented as uniquely mysterious properties of deep neural networks are neither unique to neural networks nor outside classical generalization theory:

1. benign overfitting — a model can fit training noise and still generalize on structured data;
2. overparametrization — a model can have more parameters than observations without necessarily overfitting;
3. double descent — test error can improve again after the model crosses the interpolation threshold.

The unifying idea is soft inductive bias:

Keep a flexible hypothesis space, but give higher preference to solutions that fit the data with a simpler, more compressible explanation.

This differs from a hard restriction that removes complex hypotheses from the model class. A soft bias leaves those hypotheses available when the data requires them, but makes them less preferred when a simpler explanation fits just as well.

flowchart LR
  H["Flexible hypothesis space"] --> F["Can fit structured data or noise"]
  B["Soft preference for simple / compressible solutions"] --> S["Select among equally good fits"]
  F --> S
  S --> G["Simple fit on structured data"]
  S --> N["Complex fit only when the data requires it"]

The Simplest Example

The paper uses a high-order polynomial as the main intuition pump. Instead of fixing a low polynomial degree, it allows a large degree $J$ but penalizes higher-order coefficients more strongly:

$$\mathcal{L}(w)=-\log p(y\mid f(x,w))+\sum _j{\gamma }^j{w}_j^2,f(x,w)=\sum _{j=0}^J{w}_j{x}^j,\gamma >1.$$

This model is highly expressive: it can use high-order terms and even interpolate noise. But on simple structured data it prefers lower-order coefficients because they are cheaper under the regularizer. Therefore:

• the model’s capacity to fit noise does not imply that it will choose a noisy function on structured data;
• flexibility and a strong simplicity bias can coexist;
• parameter count or maximum polynomial degree is a poor proxy for the complexity of the selected function.

The paper reports a 150th-order version of this example and compares low-order, high-order, and order-dependently regularized polynomials across multiple data sizes and target functions.

Formal Generalization Lens

For a countable hypothesis space with a data-independent prior $P(h)$, bounded loss range $\Delta$, $n$ independent observations, and confidence parameter $\delta$, the paper uses the bound

$$R(h)\le \hat{R}(h)+\Delta \sqrt{\frac{\log \frac{1}{P(h)}+\log \frac{1}{\delta }}{2n}}.$$

In plain language:

$$\text{test risk}\lesssim \text{training risk}+\text{description complexity penalty}.$$

The bound does not penalize every hypothesis merely because the overall model class is large. It penalizes the particular selected hypothesis according to how unlikely or complex it is under the prior.

With a Solomonoff-style prior $P(h)\propto 2^{-K(h\mid A)}$, where $K(h\mid A)$ is the Kolmogorov complexity of the selected hypothesis conditioned on fixed architecture information $A$, the penalty becomes a description-length term. The shortest description is not computable, but the stored model and decoding procedure can upper-bound it.

PAC-Bayes gives a related distributional view. Instead of scoring one selected hypothesis, it compares a posterior-like distribution $Q$ over useful solutions with a prior $P$ through $\mathrm{KL}(Q\Vert P)$. The paper emphasizes that deterministic models are not excluded: point-mass or countable-hypothesis variants recover closely related bounds.

How The Paper Reads The Three Phenomena

Phenomenon	Misleading interpretation	Paper’s interpretation
Benign overfitting	If a model can fit random labels, it should overfit ordinary data.	Capacity to fit noise only shows flexibility. A simplicity bias can still select a compressible structured solution on ordinary data.
Overparametrization	More parameters necessarily mean a more complex learned function.	Parameter count measures representation capacity poorly. Larger models can contain more solutions while also making flat or compressible solutions easier to represent and find.
Double descent	Once extra capacity starts overfitting, still more capacity should only hurt.	Near interpolation, effective complexity can peak; farther into the overparametrized regime, flat and compressible interpolating solutions can become more prevalent, reducing effective dimensionality and test error.

Benign overfitting

The paper distinguishes being able to fit noise from preferring a noisy solution when structure is present. The high-order polynomial provides the simplest example. It can fit pure noise, yet its order-dependent penalty makes simple structured functions preferred on structured data.

The paper also points to Gaussian-process and Bayesian-neural-network results that reproduce the noisy-label CIFAR behavior often used to motivate the claim that deep learning requires a wholly new generalization theory. Marginal likelihood falls as labels are randomized, and the paper connects marginal likelihood to PAC-Bayes-style evidence.

Overparametrization

The paper argues that parameter count is not function complexity. Infinite-basis Gaussian processes can generalize strongly on small datasets because the prior over functions supplies a soft simplicity bias. Likewise, a larger neural network may be more flexible while its trained solution has lower effective dimensionality or stronger compressibility.

The proposed intuition is geometric: in a larger parameter space, flat families of low-loss, compressible solutions can occupy much more relative volume than isolated bad fits. The paper treats this as a promising explanation rather than a completed universal mechanism.

Double descent

The paper reproduces double descent with both a ResNet and a linear random-feature model. Its preferred diagnostic is effective dimensionality, not raw parameter count. Near the interpolation threshold, effective dimensionality and test error rise; deeper in the overparametrized regime, training loss is already approximately zero, while effective dimensionality and test error can fall.

For the linear case, minimum-norm interpolation and random-matrix arguments provide a classical explanation. For neural networks, flatness, compressibility, and the volume of connected low-loss solutions supply the corresponding interpretation, with the caveat that no one scalar diagnostic fully explains generalization.

What May Actually Be Distinctive About Deep Learning

The paper does not claim that deep learning is fully understood or ordinary in every respect. It highlights three properties as more genuinely distinctive:

1. Representation learning. Neural networks learn adaptive basis functions and therefore learn a task-dependent similarity metric, rather than using only a fixed basis or fixed kernel.
2. Relative universality. The same broad architecture family can work across text, images, audio, time series, and other domains; transformers can also perform broad in-context learning without parameter updates.
3. Mode connectivity. Independently trained neural-network solutions can often be connected by low-loss paths, undermining the picture of isolated local optima and motivating methods such as stochastic weight averaging, model soups, and model merging.

The paper is careful that none of these properties is absolutely unique to neural networks. Its claim is relative: deep networks exhibit them unusually efficiently, broadly, or prominently.

Evidence Scope

The paper combines new explanatory examples with prior results:

• a high-order polynomial with order-dependent regularization demonstrates flexibility plus a soft simplicity bias;
• a Gaussian process reproduces mixed clean/noisy-label CIFAR-10 behavior;
• ResNet-18 and linear random-feature examples display double descent alongside effective-dimensionality changes;
• prior PAC-Bayes and countable-hypothesis work supplies non-vacuous bounds for large neural networks and language models;
• mode-connectivity, representation-learning, and compression-bias evidence is synthesized from cited work.

Many figures are adapted from prior papers, and all figures except the explicitly described experiments are conceptual. This is therefore stronger as a unifying interpretation and literature-bridging position than as a standalone experimental benchmark paper.

Author Narrative Context

The official 12-post X thread presents the same thesis in a compressed form: soft inductive biases allow a maximally flexible hypothesis space plus a compression preference; the polynomial and Gaussian-process examples show that benign overfitting is not neural-network-specific; and the older countable-hypothesis framework can describe the behavior.

Wilson’s official talks page makes the broader design philosophy explicit: flexibility and simplicity need not trade off, and the usual phrase “bias—variance trade-off” can be misleading when an expressive model also has a strong simplicity bias. The paper itself supports the narrower claim that low bias and low variance can coexist; it does not establish that every large model, prior, or training recipe achieves this favorable combination.

Limitations And Gotchas

• It is a position paper. It advances a selective unifying thesis rather than a comprehensive survey or a new complete theory of neural-network training.
• Reproducing a phenomenon is not a full causal explanation. Showing benign overfitting or double descent in a polynomial, Gaussian process, or linear model refutes neural-network uniqueness, but does not prove that every neural network exhibits the same mechanism.
• A valid bound can still be loose. PAC-Bayes or countable-hypothesis validity does not mean the bound predicts exact test error, model ranking, or training dynamics.
• The prior and code matter. A convenient model filesize is an upper bound on description length, not the incomputable shortest program itself. Different coding schemes can produce different practical tightness, and the prior must respect the bound’s data-dependence rules.
• The elementary bound has assumptions. The stated theorem uses bounded risk and independent observations. Extensions to language models, unbounded token losses, stochastic training, and dependent tokens require additional machinery from the cited work.
• “Larger can generalize better” is not “larger is always better.” Compute, data, optimization, latency, calibration, and deployment constraints still matter. The paper explicitly acknowledges computational benefits of restriction biases.
• Hard constraints can encode guarantees. Exact conservation laws, causal structure, safety invariants, or known control constraints may deserve hard enforcement even when soft biases are useful for approximate symmetries.
• Compression can erase what matters. In time-series and world-model settings, a compact solution may discard rare regimes, small numeric deviations, delayed effects, event timing, exogenous variables, or action consequences while still looking simple under a generic code.
• Optimizer effects remain open. The paper argues that stochastic-optimizer bias is not necessary for respectable generalization, not that optimizer choice is irrelevant under finite compute.

Foundation TSFM Relevance

Agenda slot	Verdict	Evidence	Missing pieces
Scaling and efficiency	adjacent	Separates raw parameter count from selected-function complexity, effective dimensionality, and compressibility.	Needs direct multivariate time-series scaling experiments under matched data, FLOPs, wall clock, and serving constraints.
Training dynamics	adjacent	Frames architecture, prior, regularization, and optimizer as possible soft biases that select among many interpolating solutions.	Does not test current TSFM objectives, optimizers, long horizons, irregular sampling, or action-conditioned rollouts.
Benchmark hygiene	warning	Ability to fit randomized labels or reach zero training loss does not by itself show poor generalization; the selected solution and complexity measure must distinguish structured from randomized data.	Needs time-series randomization controls that preserve autocorrelation, channel structure, event timing, and leakage boundaries.
Latent-state and representation learning	adjacent	Identifies learned adaptive representations and similarity metrics as a genuinely important deep-learning capability.	Does not test whether TSFM representations preserve regimes, constraints, event streams, control inputs, or intervention effects.
Native multivariate and control interfaces	insufficient evidence	The framework can evaluate any learned hypothesis in principle.	No direct evidence for high-channel multivariate time series, exogenous variables, actions, interventions, or counterfactual state prediction.

Because all direct time-series relevance is adjacent, warning, or insufficient evidence, this source should not be promoted into the central agenda coverage matrix.

Practical Reading Frame For Time-Series Models

When applying the paper’s lens to a time-series foundation model, separate four questions:

1. Capacity: What functions and trajectories can the architecture represent?
2. Selection: Which solution does the objective, prior, architecture, and optimizer prefer?
3. Compression: Is the selected representation simple because it captures reusable structure, or because it erased rare but decision-relevant state?
4. Generalization: Does the same solution preserve calibration, rare regimes, channel coupling, context, exogenous variables, and action effects on held-out systems?

A useful evaluation should therefore report parameter count separately from effective rank/dimensionality, compressibility, perturbation stability, train fit, structured test performance, randomized-label controls, and rare-state preservation. Compression is evidence only when the compressed state retains the capabilities the time-series or world-model task actually requires.

Links Into The Wiki

• Training Dynamics
• Time-Series Scaling And Efficiency
• Time-Series Benchmark Hygiene
• Self-Supervised Representation Learning
• Intermediate-Layer Representations
• Foundation Time-Series Model Research Agenda

Open Questions

• Do larger time-series foundation models actually develop stronger compression biases at matched data and compute, or do they mainly gain raw capacity?
• Which compression or effective-dimensionality diagnostics predict zero-shot transfer, calibration, rare-regime recall, and long-horizon state retention?
• Can a PAC-Bayes or countable-hypothesis bound rank TSFM checkpoints better than validation loss alone?
• How should conditional description length account for tokenization, sampling rate, channel count, architecture, and known physical structure?
• When should approximate temporal or spatial invariances be soft preferences, and when should physical or safety constraints remain hard?
• Does a double-descent regime appear in multivariate time-series pretraining when model size, dataset size, horizon, channel count, or context length crosses an interpolation threshold?
• Can a model become more compressible while becoming worse at rare events, interventions, or action-conditioned rollout?