Speaker Notes

The Real Reason Huge AI Models Actually Work [Prof. Andrew Wilson]

Massive deep learning models succeed by leveraging implicit simplicity biases and volume geometry that naturally favor generalizable solutions over overfitting.

Machine Learning Street Talk · 2:03:48 · YouTube ↗

Interview · 2 hosts + guest  Intermediate · Concepts and theory of deep learning generalization

Highlights

12 clips · 18 min of 124 min
    • Massive models can exhibit stronger simplicity biases than smaller ones.
    • The classical bias-variance trade-off does not always hold in modern deep learning.
    • Parameter counts fail to accurately reflect true model complexity.
    • The induced distribution over functions dictates model behavior and preferences.
    • Increasing model flexibility first hurts and then helps generalization through double descent.
    • Massive parameter scales naturally lean toward simple solutions consistent with observations.
    • Building larger models can alleviate overfitting past the interpolation threshold.
    • Once training loss hits zero, other biases like solution flatness dominate model selection.
    • The empirical prescription is to make models as large as possible combined with a compression bias.
    • Deep ensembles approximate the Bayesian predictive distribution better than traditional MCMC or laplace approximations.
    • Empirically successful non-Bayesian methods often actually represent better Bayesian marginalization.
    • Deep ensembles were mistakenly categorized as a non-Bayesian alternative due to performance comparisons.
    • For a given computational budget, ensembles do a superior job of approximating the Bayesian predictive distribution.
    • The correct historical correction is that model construction needs to be more Bayesian, not less.
    • Soft biases incur a small penalty for deviating from a structural constraint like equivariance.
    • If the constraint fits the data, the model naturally collapses down onto that constrained solution.
    • Gentle priors achieve comparable performance to perfectly constrained models while retaining flexibility.
    • Full-batch gradient descent achieves generalization similar to SGD without explicit implicit regularization tricks.
    • Even random sampling (guess-and-check) yields comparable generalization in heavily over-parameterized models.
    • The geometric volume of low-loss regions drives generalization rather than optimizer dynamics alone.
    • Training transformers exclusively on Gaussian random matrices fails when tested on out-of-distribution matrices like identity matrices.
    • Enriching the training distribution with diverse structured matrices enables the transformer to learn underlying algorithms rather than just interpolating.
    • Data-driven algorithm discovery is viable when training sets cover a sufficiently rich space of structures.
    • Text-based pre-training predisposes models to discover compressible representations and salient physical symmetries like rotation invariance.
    • Compression bias acts as a universal induction principle across vastly different low-level modalities.
    • Vision transformers can achieve lower translation equivariance error than purpose-built convolutional neural networks at scale.
    • More complex models can generate a wider variety of datasets, forcing them to spread their probability mass more thinly.
    • Marginal likelihood automatically penalizes unnecessary model complexity by measuring the volume of supported data distributions.
    • This formalizes Occam's razor mathematically without requiring manual regularization hacks.
    • Combinatorial optimization like integer programming is difficult due to a lack of smoothness.
    • Expanding the search space to floating-point numbers enables smooth gradient-based optimization.
    • Smooth navigation helps the optimizer successfully land on near-optimal discrete solutions.
    • Parameter sharing can hinder compute-optimal scaling when training data is unlimited.
    • With a negligible generalization gap, the goal shifts to maximizing parameters per flop.
    • Fine-grained mixture of experts and structured matrices improve efficiency by operating over individual layers.

Questions this video answers

14 · every answer cites the moment it was said
01 Why is the bias-variance trade-off considered a misnomer in deep learning? 2 sources · 2:13

The classical bias-variance trade-off assumes you must choose between fitting data well and stability under sampling. However, modern deep learning shows that massive models can achieve both low bias and low variance simultaneously. This happens because scale introduces an implicit simplicity bias that naturally favors generalizable solutions over overfitting.

02 Why is parameter count a poor proxy for model complexity? 2 sources · 9:01

Counting parameters fails to capture the actual properties of the induced distribution over functions. An expressive model can have many parameters while maintaining strong preferences or soft constraints for certain types of solutions. Therefore, flexibility and simplicity bias can coexist, allowing models to adapt properly across data regimes.

03 How do soft constraints compare to hard constraints in model construction? 2 sources · 9:36

Hard constraints rigidly restrict the hypothesis space, which often misrepresents real-world complexities and fails if assumptions are slightly off. Soft constraints keep the model expressive while penalizing deviations from desired behaviors. When consistent with observations, expressive models naturally collapse down onto these soft constraints.

04 Why are deep ensembles better than approximate Bayesian methods? 2 sources · 31:11

For a given computational budget, deep ensembles do a much better job of approximating the true Bayesian predictive distribution and marginalization. Historical comparisons incorrectly framed them as non-Bayesian alternatives when they actually act more Bayesian.

05 How do soft constraints compare to hard constraints in model construction? 2 sources · 33:15

Soft biases or gentle encouragements toward certain constraints can perform just as well as perfectly constrained models. Because they pay a small penalty for deviating, they collapse onto the constraint if it fits the data well.

06 What drives the simplicity bias in large neural networks? 2 sources · 50:42

Geometric properties of the loss landscape, particularly the exponential dominance of the volume of flat solutions as parameter count grows, drive simplicity biases. Optimization methods like SGD or even random search land in these generalizable flat regions.

07 How well do transformers handle matrix operations without specialized tools? 3 sources · 1:03:21

Transformers can handle in-distribution matrix operations reasonably well when trained on standard normal matrices, but they fail completely outside that distribution. Enriching the training distribution with diverse matrix structures allows the model to generalize and discover underlying algorithms.

08 Can text pre-trained LLMs be applied effectively to time series or materials generation? 2 sources · 1:06:11

Yes, off-the-shelf text-trained LLMs can perform surprisingly well on time series forecasting and materials generation tasks. When fine-tuned on atomistic data, they can generate inorganic crystals with favorable properties better than purpose-built models.

09 Why is Bayesian marginalization important for highly expressive deep learning models? 2 sources · 1:13:46

Marginalization represents honest uncertainty over what solution is correct given finite data, preventing the model from betting everything on a single parameter setting. As model expressivity increases, more parameter settings fit the observed data, making marginalization crucial for generalization.

10 How does stochastic optimization relate to flat minima and Bayesian methods? 2 sources · 1:22:49

Stochastic optimization with small batches tends to find flat minima rather than sharp ones, resulting in better test loss and generalization. Bayesian marginalization naturally places most of its volume in these flat regions, offering an elegant alternative to manually defining and penalizing sharpness in loss functions.

11 What is mode connectivity in neural networks? 2 sources · 1:39:48

Mode connectivity demonstrates that separately initialized neural network solutions are not isolated by high-loss barriers. Using simple parametric curves like quadratic Bezier curves, you can walk between different modes without increasing the training loss along the path.

12 How does increasing model size affect simplicity bias? 2 sources · 1:47:26

Making neural network models larger often enhances rather than reduces a natural simplicity bias. This explains why overparameterized models generalize better in the second descent phase despite having zero training loss.

13 Why did parameter sharing fail to improve compute-optimal scaling? 2 sources · 1:55:44

Parameter sharing like in convolutions turned out not to be optimal when there is an abundance of data and no generalization gap. In that data-rich regime, the optimal strategy is maximizing parameters per flop rather than sharing parameters.

14 What is the ultimate goal for advanced AI systems beyond current applications? 2 sources · 2:00:35

The ultimate goal is developing AI systems capable of discovering fundamental scientific theories like general relativity or quantum mechanics. Moving beyond black-box function approximators will yield genuine new insights into physics and the natural world.

Section-by-section summary

24 parts

Deep learning's universality and the bias-variance misnomer

Deep learning is distinguished primarily by its broad applicability and representation learning capabilities rather than inherent mysteries. The classical bias-variance trade-off fails to capture how modern overparameterized models can achieve both low bias and low variance.

Challenging misconceptions about scale and simplicity

Research fields often get stuck in local minima due to strongly held misconceptions regarding model construction and stochastic optimization. Scale introduces a powerful simplicity bias that helps models generalize even with limited data points.

Parameter count is a poor complexity proxy

Counting parameters is a poor proxy for model complexity compared to evaluating the induced distribution over functions. Expressive models combined with strong simplicity biases automatically adapt well to both small and large datasets.

Combining scientific understanding with engineering details

Pairing classical theory with empirical observations yields durable understanding that outlives specific algorithms. Scientists use control to gain knowledge while engineers use knowledge to gain control.

Soft constraints and double descent explain scale

Rather than rigid geometric constraints, soft constraints and high expressiveness let models collapse onto desirable solutions naturally. Double descent demonstrates that larger models alleviate overfitting through emerging simplicity biases.

The bias-variance trade-off is a misnomer

Ensembling and large neural networks achieve both low bias and low variance, directly disproving the traditional trade-off. Flatness is just one of many properties controlling generalization, and low loss remains the primary goal.

Real-world data distribution and Bayesian deep ensembles

Real-world data occupies a small corner of possible problems, heavily biased toward local complexity. Deep ensembles actually provide a superior approximation of the Bayesian predictive distribution compared to methods traditionally labeled as Bayesian.

Deep ensembles actually approximate Bayesian inference better

Early misconceptions framed deep ensembles as a non-Bayesian alternative because they outperformed approximate Bayesian methods. In reality, deep ensembles do a superior job of marginalization, proving we need to be more Bayesian.

Soft biases match hard constraints with better expressiveness

Enforcing structural constraints via soft penalties—like residual pathway priors for equivariance—allows models to maintain expressiveness while collapsing onto the constraint if it fits the data. Gentle parameter priors prevent extreme outcomes while preserving flexibility.

Expressive models can still possess strong simplicity biases

Infinitely wide models like Gaussian processes demonstrate that extreme expressiveness can coexist with powerful simplicity preferences, leading to high data efficiency. We can reconcile Radford Neil's thesis on complexity with overfitting by separating model expressiveness from complexity.

Large models implicitly approximate Solomonov induction

Double descent reveals that over-parameterized models generalize well because they are implicitly biased toward low Kolmogorov complexity solutions. Tight generalization bounds are achievable using Solomonov priors, proving that scale enhances simplicity bias.

Loss landscape geometry drives the implicit simplicity bias

The volume of flat solutions grows exponentially with model size, causing optimization to naturally favor compressible parameters regardless of whether full-batch or SGD is used. This geometric shift explains why larger models generalize better without hand-crafted regularizers.

Transformers struggle with out-of-distribution primitive operations

While large models capture low Kolmogorov complexity patterns, training on Gaussian random matrices leaves transformers brittle when encountering out-of-distribution structures like identity matrices. True general intelligence requires robust internal representations of core primitives rather than relying solely on shortcuts.

Learning fundamental primitives inside transformers

While LLMs can use external tools, training them on fundamental primitives like matrix operations helps build robust internal representations. Exposing models to a rich distribution of structured matrices enables true algorithmic generalization beyond simple interpolation.

Text pre-training transfers to time series and materials

Off-the-shelf text LLMs applied naively to time series forecasting and inorganic crystal generation perform surprisingly well. Text-based pre-training instills a compression bias that helps models quickly discover necessary invariances and symmetries.

Representing uncertainty through Bayesian marginalization

Being Bayesian means maintaining an honest representation of epistemic uncertainty over solutions given finite data. Marginalizing over parameter space becomes increasingly crucial as model expressivity and parameter counts grow.

Connecting Bayesian methods to flat minima

Stochastic optimization naturally finds flat minima that generalize better, aligning with Bayesian marginalization which naturally weighs flat regions more heavily. Although full Bayesian inference is computationally challenging, approximate methods like Gaussians yield clear performance gains.

Using marginal likelihood for scientific hypothesis testing

Marginal likelihood provides a principled way to perform model selection and scientific hypothesis testing by quantifying how well a model predicts observed data without overfitting. This approach successfully distinguishes general relativity from alternative theories by penalizing overly flexible hypotheses.

Training dynamics link compression to model quality

Viewing model training through information theory connects marginal likelihood to sequential coding and data compressibility. Models typically learn structure before noise, reflecting an evolution driven by compute-limited complexity.

Scaling laws and grocking stem from representational reorganization

Phenomena like grocking and double descent occur because continued training forces models to reorganize into simpler, highly compressible representations. Stochastic weight averaging helps exploit these flat loss regions to achieve better generalization.

Mode connectivity reveals continuous flat manifolds

Retraining neural networks from different initializations demonstrates that distinct local optima are often connected by flat paths with zero loss barrier. This geometry indicates that loss landscapes form high-dimensional manifolds rather than isolated pockets.

Balancing expressiveness and simplicity bias

Effective model design requires combining expressive architectures with a robust simplicity bias, which larger models naturally exhibit through scaling. Researchers aim to engineer these compression biases into smaller models directly.

Inductive biases can improve scaling exponents

While Rich Sutton's bitter lesson emphasizes brute-force compute over hand-crafted features, learning fundamentally requires making assumptions. Designing structured layers like block tensor trains can actively improve scaling exponents and data efficiency.

The ultimate goal is discovering fundamental scientific theories

Current large language models excel at benchmark tasks but lack genuine scientific insight or understanding. The next frontier is automating the discovery of fundamental physical and biological theories rather than relying on black-box curve fitting.

0:00 / 2:03:48
k play/pause · f fullscreen · m mute
n next · p previous clip · t highlights/full
< > slower/faster · / search what was said
  • Simplicity bias22× · first 6:00

    The inherent tendency of models, particularly large ones, to favor simpler and more generalizable solutions.

  • Double descent14× · first 1:46

    A phenomenon where test error decreases, increases, and then decreases again as model capacity grows.

  • Solomonov prior2× · first 45:15

    A prior assigning exponentially stronger preferences to solutions with low Kolmogorov complexity.

  • Kolmogorov complexity6× · first 45:34

    The shortest possible program that can generate a given hypothesis or data sequence.

  • AAM's razor15× · first 1:37

    The induction principle that simpler explanations or compressed representations are generally preferred and lead to better generalization.

  • Marginal likelihood11× · first 1:25:04

    The probability of generating a particular dataset under a model prior, acting as a built-in penalty for complexity in Bayesian model selection.

  • Mode connectivity4× · first 0:59

    The discovery that different local optima in a neural network loss landscape can be connected by paths of zero training loss.