Microscopic and collective signatures of feature learning in neural networks
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.
Kai: I'm Kai.
Mira: And I'm Mira, here with Lev.
Kai: We're looking at the paper: "Microscopic and collective signatures of feature learning in neural networks".
Mira: Here's the gist of the paper. Microscopic and collective signatures of feature learning in neural networks investigates analytical fingerprints of feature learning in Bayesian one-hidden-layer neural networks by quantifying collective and microscopic observables in the proportional limit The posterior average squared distance is a nonmonotonic function of the Gibbs temperature, which is interpreted as the equilibrium counterpart of a phenomenon observed under gradient descent dynamics Collective Observables and Feature Segregation The study focuses on computing the posterior average squared distance, denoted as ⟨D2⟩, between class manifolds in a binary classification problem within Bayesian one-hidden-layer (1HL) FC NNs The authors show that this observable is a nonmonotonic function of the Gibbs temperature T This nonmonotonicity is a signature of feature learning at proportional width Specifically, the distance undergoes a finite shift in the proportional regime due to microscopic weight displacements The squared distance D2 is defined as the separation between the mean post-activations of the two different classes Microscopic Weight Correlations The analysis extends to microscopic features by investigating the statistics of the first-layer weights w, showing that these statistics depend on the data in a way that is absent in the infinite-width limit The two-point function ⟨w1kw1h⟩ with respect to the posterior distribution is expressed using an effective action formalism The resulting expression for ⟨W1hW1k⟩ in the proportional limit includes terms dependent on data-dependent kernels Λ1 and Λ2 The second correction term, Λ2, contributes to the statistics of the weights with a finite term that depends almost linearly on α The Role of Temperature and Training Dynamics The temperature T acts as a regularizer, playing the role of early stopping in optimization dynamics The behavior of the distance aligns with inversion dynamics observed during training The peak at the inversion point corresponds to typical loss values associated with the transition between easy and hard samples Increasing temperature allows the posterior to sample regions of the loss landscape further away from optima Theoretical Framework and Limits The analysis is conducted within a statistical-mechanics framework for Bayesian learning in one hidden layer neural networks with standard parameterization The results are computed analytically in the proportional limit, where the partition function can be evaluated via saddle-point evaluation of an effective action S(Q) The second-order statistics of the hidden-layer weights develop finite correlations in the proportional limit, which is remarkable In the infinite-width limit, these terms vanish, indicating that no feature learning happens The results hold for standard-scaled neural networks and represent a genuinely overparameterized scenario The squared distance D2 is identified as a collective observable of the features that signals nontrivial feature learning in Bayesian FC shallow networks The behavior of the distance reflects a trade-off between the segregation of two class manifolds and the fine tuning required for classifying hard samples The posterior second-order statistics of the hiddenlayer weights develop finite correlations in the proportional limit The temperature at the inversion peak corresponds to typical loss values associated with the transition between easy and hard samples Derivation Details The derivation of the average squared distance involves decomposing D2 into three contributions: D˜ 2 1, D˜ 2 2, and D˜ 2 3 The term D˜ 2 3 is proportional to the posterior expected output difference squared The zero-temperature limit can be used to justify different scalings in P of the three quantities The second contribution D˜ 2 2 can be neglected under certain assumptions, and the first term D˜ 2 1 is of the same order as D˜ 2 3 The scaling argument allows considering these two contributions effectively of the same order Consistency Checks A consistency check for a linear activation function shows that the resulting expectation value matches the one obtained by an independent computation The expression for ⟨W1hW1k⟩ in Eq.
Kai: (52) is equivalent to Eq.
Mira: (13) in the case of a linear activation function The kernel Kµν = Cµν, and the final result matches the independent computation The terms involved return Cµν, -CµλCνσ, and terms involving Q¯ which are consistent with Eq.
Paper summary: Kai: So, looking at the "Microscopic and collective signatures of feature learning in neural networks" paper by Corti and Pacelli and Gherardi, they are suggesting that the average squared distance is a collective observable that signals nontrivial feature learning in Bayesian FC shallow networks.
Mira: They connect this to the way the distance changes with temperature T, showing it's nonmonotonic, which reflects a trade-off between separating class manifolds and fine tuning for hard samples.
Lev: It seems like they’ve pinpointed exactly where we should focus our attention when thinking about feature learning in these models. The implication is that this specific behavior at proportional width is what separates actual learning from just having a lot of parameters.
Kai: They found that the posterior second-order statistics of the hidden layer weights develop finite correlations in the proportional limit, which is a significant finding because it shows something non-trivial happening when you can’t take the infinite width limit.
Mira: The temperature at the inversion peak corresponds to typical loss values associated with that transition between easy and hard samples, giving us a way to characterize what's happening during training dynamics.
Lev: For someone listening who doesn't know all this deep math, this suggests that feature learning isn't just about finding features; it’s also about the specific geometry of how those features are organized and how they adapt based on the data scale.
Conclusion: Kai: So we're wrapping up on this paper about microscopic and collective signatures of feature learning in neural networks.
Mira: Basically, they’re saying that when you look at a one hidden layer network, there are measurable fingerprints—collective observables—that tell you if the network is actually learning features or just memorizing things.
Kai: It's about checking how those fingerprints change depending on the temperature setting and the width of the network.
Mira: They found this average squared distance, which is a collective observable, isn't just a flat number; it goes up and down with the Gibbs temperature. That nonmonotonic behavior is what they call a signature of feature learning at proportional width.
Kai: So, it’s like there’s this specific sweet spot where the network starts organizing its features in a meaningful way, but if you push the width too far, that organization breaks down.
Mira: Exactly. And they also looked at the actual weights inside the hidden layer and found that their statistics develop these finite correlations in a limit where we usually expect them to vanish.
Kai: That’s interesting because it suggests that even in really overparameterized models, there's still some fine-grained structure happening with those weights.
Mira: It points to a trade-off, something between the network successfully grouping the data and having enough precision left for classifying those trickier samples.
Kai: It means feature learning isn't just one single thing; it’s this complex interplay between how features are segregated and how finely they are tuned.
A. Corti, R. Pacelli, P. Rotondo, M. Gherardi
Universita degli Studi di Milano · I.N.F.N. · Dipartimento di Scienze Matematiche, Fisiche e Informatiche, Universita degli Studi di Parma
cond-mat.dis-nn, cond-mat.stat-mech
Submitted: 2025-08-28
Updated: 2026-11-16
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 97/100
The gist: Microscopic and collective signatures of feature learning in neural networks investigates analytical fingerprints of feature learning in Bayesian one-hidden-layer neural networks by quantifying…
Key concepts
- Posterior Average Squared Distance (⟨D2⟩)
- This observable measures the separation between two classes by looking at the mean post-activations of those classes. Its nonmonotonic change with temperature is a key signature indicating how features are learning in a neural network.
- Gibbs Temperature (T)
- Temperature acts as a regularizer, mimicking early stopping during training dynamics. The paper shows that the distance metric's behavior at different temperatures reflects the transition between easy and hard samples in the loss landscape.
- Proportional Limit
- This limit is used for analytical calculations where the network width is considered proportional to its number of parameters. In this regime, second-order statistics of hidden weights develop finite correlations, which indicates feature learning occurs.
- Microscopic Weight Correlations
- The analysis examines the statistical properties of individual first-layer weights (w). These statistics depend on the data in a way that is absent in the infinite-width limit, proving that specific weight patterns are learned during training.
Terminology
Summary
Microscopic and collective signatures of feature learning in neural networks investigates analytical fingerprints of feature learning in Bayesian one-hidden-layer neural networks by quantifying collective and microscopic observables in the proportional limit The posterior average squared distance is a nonmonotonic function of the Gibbs temperature, which is interpreted as the equilibrium counterpart of a phenomenon observed under gradient descent dynamics
Collective Observables and Feature Segregation
The study focuses on computing the posterior average squared distance, denoted as ⟨D2⟩, between class manifolds in a binary classification problem within Bayesian one-hidden-layer (1HL) FC NNs The authors show that this observable is a nonmonotonic function of the Gibbs temperature T This nonmonotonicity is a signature of feature learning at proportional width Specifically, the distance undergoes a finite shift in the proportional regime due to microscopic weight displacements The squared distance D2 is defined as the separation between the mean post-activations of the two different classes
Microscopic Weight Correlations
The analysis extends to microscopic features by investigating the statistics of the first-layer weights w, showing that these statistics depend on the data in a way that is absent in the infinite-width limit The two-point function ⟨w1kw1h⟩ with respect to the posterior distribution is expressed using an effective action formalism The resulting expression for ⟨W1hW1k⟩ in the proportional limit includes terms dependent on data-dependent kernels Λ1 and Λ2 The second correction term, Λ2, contributes to the statistics of the weights with a finite term that depends almost linearly on α
The Role of Temperature and Training Dynamics
The temperature T acts as a regularizer, playing the role of early stopping in optimization dynamics The behavior of the distance aligns with inversion dynamics observed during training The peak at the inversion point corresponds to typical loss values associated with the transition between easy and hard samples Increasing temperature allows the posterior to sample regions of the loss landscape further away from optima
Theoretical Framework and Limits
The analysis is conducted within a statistical-mechanics framework for Bayesian learning in one hidden layer neural networks with standard parameterization The results are computed analytically in the proportional limit, where the partition function can be evaluated via saddle-point evaluation of an effective action S(Q) The second-order statistics of the hidden-layer weights develop finite correlations in the proportional limit, which is remarkable In the infinite-width limit, these terms vanish, indicating that no feature learning happens The results hold for standard-scaled neural networks and represent a genuinely overparameterized scenario
Conclusion on Feature Learning
The squared distance D2 is identified as a collective observable of the features that signals nontrivial feature learning in Bayesian FC shallow networks The behavior of the distance reflects a trade-off between the segregation of two class manifolds and the fine tuning required for classifying hard samples The posterior second-order statistics of the hiddenlayer weights develop finite correlations in the proportional limit The temperature at the inversion peak corresponds to typical loss values associated with the transition between easy and hard samples
Derivation Details
The derivation of the average squared distance involves decomposing D2 into three contributions: D˜ 2 1, D˜ 2 2, and D˜ 2 3 The term D˜ 2 3 is proportional to the posterior expected output difference squared The zero-temperature limit can be used to justify different scalings in P of the three quantities The second contribution D˜ 2 2 can be neglected under certain assumptions, and the first term D˜ 2 1 is of the same order as D˜ 2 3 The scaling argument allows considering these two contributions effectively of the same order
Consistency Checks
A consistency check for a linear activation function shows that the resulting expectation value matches the one obtained by an independent computation The expression for ⟨W1hW1k⟩ in Eq. (52) is equivalent to Eq. p. 0 Recent progress on the theoretical analysis of NNs
has established an equivalence between standard-scaled
Bayesian FC NNs and kernel regressions, particularly in
two limits. (i) In the infinite-width framework, where the
width of the hidden layers N1 is much larger than the
number of training examples P, standard-scaled NNs are
known to be mathematically equivalent to fixed Gaussian Processes (GPs), which only depend on the prior
statistics of the weights [9–17]. In the proportional case, since the kernel ∆µν
hk cannot be expressed
in simple terms by using Kµν, the zero-temperature limit
is not useful to investigate the magnitude of the proportional correction terms. 2, we show results
of numerical simulations performed with a linear activation function.Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Collective feature learning signature detection: The system can monitor class manifold separation as a nonmonotonic function of temperature, which is interpreted as the
equilibrium counterpart of a phenomenon observed under gradient descent (GD) dynamics.
This allows for identifying training regimes where features transition between well-separated and approaching states. -
Microscopic weight correlation analysis: The system can quantify
microscopic pairwise correlations of the weights
by computing the two-point function ⟨w1kw1h⟩ with respect to the posterior distribution, revealing thatthe statistics of the first-layer weights w depends on the data (which does not happen in the infinite-width limit).
-
Feature learning phase identification: The system can determine if a network is in an
easy and hard phase of training
by analyzing whether the squared distance D2 exhibits a transition, which is interpreted asa trade-off between the segregation of the two class manifolds and the fine tuning required for classifying the last hard samples.
-
Overparameterized regime characterization: The system can distinguish between different learning regimes in overparameterized networks by noting that
the proportional regime represents a genuinely overparameterized scenario,
allowing it to characterize feature learningcloser to the interpolation threshold.
-
Finite-width weight dependency measurement: The system can measure the correlation of hidden layer weights, showing that
the second-order statistics of w depends on the input patterns in a way that cannot be traced back to the infinite-width limit,
providing a more detailed understanding than fixed Gaussian Process descriptions.
Abstract
Feature extraction - the ability to identify relevant properties of data - is a key factor underlying the success of deep learning. Yet, it has proved difficult to elucidate its nature within existing predictive theories, to the extent that there is no consensus on the very definition of feature learning. A promising hint in this direction comes from previous phenomenological observations of quasi-universal aspects in the training dynamics of neural networks, displayed by simple properties of feature geometry. We address this problem within a statistical-mechanics framework for Bayesian learning in one hidden layer neural networks with standard parameterization. Analytical computations in the proportional limit (when both the network width and the size of the training set are large) can quantify fingerprints of feature learning, both collective ones (related to manifold geometry) and microscopic ones (related to the weights). In particular, (i) the distance between different class manifolds in feature space is a nonmonotonic function of the temperature, which we interpret as the equilibrium counterpart of a phenomenon observed under gradient descent (GD) dynamics, and (ii) the microscopic learnable parameters in the network undergo a finite data-dependent displacement with respect to the infinite-width limit, and develop correlations. These results indicate that nontrivial feature learning is at play in a regime where the posterior predictive distribution is that of Gaussian process regression with a trivially rescaled prior.
Sources
- A theory of representation learning gives a deep generalisation of kernel methods
- Information-theoretic reduction of deep neural networks to linear models in the overparametrized proportional regime
- Order parameters and phase transitions of continual learning in deep neural networks
- Optimal generalisation and learning transition in extensive-width shallow neural networks near interpolation
- Statistical mechanics of extensive-width Bayesian neural networks near interpolation
- The Nuclear Route: Sharp Asymptotics of ERM in Overparameterized Quadratic Networks
- Applications of Statistical Field Theory in Deep Learning
Related papers
- Topological metal-insulator transitions in one-dimensional non-Hermitian quasicrystals: beyond PT-symmetry
- Synchronous and asynchronous Monte Carlo dynamics of mean-field spin glasses with rotationally invariant couplings: what they share and where they differ
- Reentrant Localization Length in a Quasiperiodic Thue-Morse Chain
- Beyond Ground States: Learning Many-Body Hamiltonians for p-Bit Invertible Arithmetic
- Linear equivalence of nonlinear recurrent neural networks
- Cup and Cap Topological Neural Network