Estimating Information Flow in Deep Neural Networks

Problem

• Binning is computationally attractive but coarse bins can produce inaccurate estimates, including apparent long-term decreases in I(X;Tℓ).As bin sizes shrink, the binned proxy approaches the true value by definition.Source: §1 · p. 1
• For deterministic DNNs with strictly monotone nonlinearities, true I(X;Tℓ) is infinite for continuous X or constant for discrete X, so observed fluctuations cannot reflect true compression.For discrete inputs, the mutual information equals H(X) under the stated injectivity condition.Source: §1 · pp. 1–2

Framework

• The proposed framework adds i.i.d. Gaussian noise to hidden-layer outputs, making X→Tℓ stochastic, isotropic, and compatible with meaningful parameter-dependent mutual information.The noisy networks preserve performance and learned representations when the noise variance is not too large.Source: §1 · p. 2

Mechanism

• A rigorous estimator tracks I(X;Tℓ) in noisy DNNs, while an AWGN-channel interpretation relates decreasing information to same-class representations becoming increasingly clustered and indistinguishable.Clusters also tighten in deeper layers, according to the supplied introduction passage.Source: §1 · p. 2

Findings

• Experiments show noisy-network compression, clustering in noisy and deterministic networks, and that prior binned-information experiments measured clustering rather than true mutual-information compression.The paper further reports that clustering is the central geometric phenomenon of interest.Source: §1 · pp. 2–3

Noisy DNN model

• Each hidden layer is modeled as Tℓ=fℓ(Tℓ−1)+Zℓ with Gaussian noise, while the final layer remains deterministic.The functions fℓ can represent fully connected, convolutional, max-pooling, and other layer types.Source: §2 · p. 3

Empirical validation

• On MNIST CNNs, internal noise with β≤0.1 minimally affects classification performance, while noisy and dropout models learn representations similar to the deterministic model.Increasing β reduces similarity, and similarity is lower in deeper internal layers.Source: §2 · p. 3

Mutual information

• The noisy network induces a joint distribution in which X−T1−…−TL forms a Markov chain, enabling mutual information analysis at each hidden layer.The mutual information is defined from the representation entropy and conditional entropy.Source: §2 · p. 3
• Because composed Gaussian noises and nonlinearities make the stochastic map analytically complicated, I(X;Tℓ) is estimated from samples.The estimator operates on the induced conditional and marginal distributions of noisy representations.Source: §2 · p. 3

3. Mutual Information Estimation

• The sample-propagation estimator decomposes I(X; Tℓ) into unconditional and conditional differential-entropy estimates derived from noisy-network samples.Each entropy is approximated using a Gaussian mixture determined by sampled representations and the noise distribution.Source: §3 · p. 3, §3.1 · pp. 3–4

3.2. Theoretical Guarantees and Computation

• The estimator reduces each entropy calculation to estimating h(pS ∗ ϕβ) from i.i.d. samples and known Gaussian noise, using Gaussian-convolution entropy theory.The resulting risk converges at the best possible parametric rate, though high-dimensional estimation can require exponentially many samples in the worst case.Source: §3.2 · p. 4
• The estimator becomes easier to use as noise increases because Gaussian smoothing simplifies the distribution, while smaller noise requires more samples for accurate estimation.The practical choice of β balances classifier performance, regularization, and estimator accuracy; theorem-sized sample requirements may be reduced after stability checks.Source: §3.2 · p. 4
• Monte Carlo integration evaluates the Gaussian-mixture entropy numerically, with an unbiased proxy whose mean-squared error is O(...) and adds negligible estimation error.The paper presents this computation as an efficient way to evaluate mixtures without closed-form entropy expressions.Source: §3.2 · p. 4
• Theorem 1 requires n = 4 · 10^9 samples, but good accuracy is achieved at n = 4 · 10^6, and adding more samples beyond this does not change the results.Source: §3.2 · p. 5

4. Compression and Clustering

• Mutual information measures how distinguishable noisy representation symbols are, so reducing spacing between constellation points reduces I(X; T).Under Gaussian noise, overlapping component distributions correspond roughly to fewer resolvable clusters.Source: §4 · p. 5
• In the tanh example, same-class Gaussian components progressively re-converge, producing mutual-information dips as individual representations merge.The dips near k = 10^3 and k = 10^4 correspond to successive Gaussian mergers.Source: §4 · p. 5
• For larger noise, Gaussian components can be indistinguishable throughout training, whereas smaller noise allows distinct components to merge in stages.The example therefore links the observed compression pattern to the interaction between representation spacing and noise level.Source: §4 · p. 5

5.1. Noisy SZT Model

• Compression in noisy SZT networks is driven by clustering of representations, with same-class samples forming tighter groups during training.In the tanh SZT model, saturation pushes representations toward cube corners; linear networks can also compress through clustering and dimensional projection.Source: §5.1 · pp. 5–6, §5.3 · p. 8
• Orthonormal-weight regularization achieved 98% test accuracy, while compression was completely gone, most neurons did not saturate, and no later-stage clustering was observed.Source: §5.1 · pp. 5–6

5.2. Deterministic MNIST CNN

• In the deterministic MNIST CNN, whole-layer entropy remains approximately ln(10000) = 9.210 across layers and training epochs, while most layer-3 units undergo compression, with the least-negative mean slope −0.002 nats/epoch.After one epoch, fully connected layers 3 and 4 separate the distributions of within-class and between-class pairwise distances.Source: §5.2 · pp. 7–8

5.3. Summary of Experiments

• Across experiments, binned entropy and mutual-information fluctuations track clustering rather than true mutual-information compression, which is impossible or vacuous in deterministic networks.The authors identify geometric clustering as the common phenomenon in noisy and deterministic networks, while binned measures remain useful for tracking clustering changes.Source: §5.1 · p. 7, §5.3 · p. 8

• The study identifies clustering of learned representations as the geometric phenomenon underlying apparent compression in deterministic networks.Source: §6 · p. 8
• Binned mutual information in prior deterministic-network experiments measured increased clustering rather than true mutual-information compression.Source: §6 · p. 8
• Mutual information tracks clustering in noisy networks but is meaningful only when the network includes a mechanism such as noise for shedding information.Source: §6 · p. 8
• Although binning is not an accurate mutual-information estimator, it remains computationally simple and useful for tracking clustering changes.Source: §6 · p. 8

• Leaky-ReLU training clusters inputs through coordinated weight and bias changes across the two layers.Source: A. Two-Neuron Leaky-ReLU Network Example · p. 9
• The first layer merges the X1/4 elements, while the second layer successively splits and coalesces them as biases and weights evolve.Source: A. Two-Neuron Leaky-ReLU Network Example · p. 9
• The mutual-information curves reflect these merging, spreading, splitting, and coalescing events, and the bounds precisely track their evolution.Source: A. Two-Neuron Leaky-ReLU Network Example · p. 9

B.1. SZT Model

• Larger noise can preserve clustering and neuron saturation while making compression very mild, because within-cluster distinctions are blurred before clusters tighten.Source: §B.1 · p. 9
• ReLU networks show almost no compression: unconstrained positive activations can spread without bound, while weight normalization bounds growth and reduces saturation.Source: §B.1 · p. 9, Figure 10 · p. 11

B.2. Spiral Model

• Spiral experiments with β = 0.005 show a connection between mutual-information estimates, discrete entropy estimates, and clustering in normalized and unnormalized models.Source: §B.2 · p. 10
• For deterministic spiral networks, the estimated entropy depends on bin size and reveals a particular clustering granularity rather than directly estimating mutual information.Source: §B.2 · p. 10

B.3. MNIST CNN

• The MNIST CNN experiments use two convolutional layers, two fully connected layers, batch normalization, and standard training with augmentation.Source: §B.3 · p. 10, §15 · p. 10

13. Tanh() activation function

• The cited architecture uses 2-D convolutional layers with 1-to-16 and 16-to-32 channels, 5x5 kernels, two-pixel padding, and either Tanh activations or dropout at probability 0.2.Source: §B.3 · p. 10

15. Fully connected layer with 128 inputs and 10 outputs

• At test time, noisy and dropout models are made deterministic by replacing noise and dropout blocks with identities, defining the reported layer outputs.Source: §15 · p. 12

C.1. Preliminary Definitions

• The estimator framework treats Gaussian-convolved hidden-layer distributions as continuous and defines minimax entropy-estimation risk over distribution classes.Source: §C.1 · p. 12

• Sample complexity for accurately estimating h(P ∗ ϕβ) is exponential in dimension d.Source: §C.2 · p. 12
• For fixed β > 0, the lower bound applies when dimension is sufficiently large and the target error is sufficiently small.Source: §C.2 · p. 12
• For every fixed dimension, sufficiently small β also yields a sample-complexity lower bound.Source: §C.2 · p. 13
• Larger β values are favorable for estimation, whereas high dimension makes exponential sample complexity unavoidable.Source: §C.2 · p. 13

• The convergence rate cannot improve beyond the parametric rate, and its exponential dependence on dimension is necessary.The stated lower bounds rule out faster dependence on sample size or dimension under the considered setup.Source: §C.3 · p. 14
• Sampling from the noisy DNN produces unconditional and conditional samples used to estimate the entropies defining mutual information.Unconditional samples are generated by propagating each input once, whereas conditional samples reuse each input across independent noise realizations.Source: §C.3 · pp. 14–15
• The estimator’s expected absolute error converges at the parametric estimation rate for any hidden-layer dimension.The guarantee follows by combining differential-entropy estimation bounds with the sample-propagation mutual-information estimator.Source: §C.3 · pp. 14–15
• Larger noise makes estimation easier because 1/β^2 is the signal-to-noise ratio between the signal and injected noise.The estimator is valid for layers ℓ≥2; for the first layer, the conditional entropy is replaced by the known entropy of the additive Gaussian noise.Source: §C.3 · p. 15
• The absolute-risk bound depends on the representation dimension rather than the ambient input dimension because additive noise acts in the representation domain.Noise smooths the conditional distributions, uniformly controls conditional-entropy variance, and leaves the input contribution as an empirical average with rate 1/√n.Source: §C.3 · p. 15

C.4. Sample Propagation Estimator Bias

• The bias analysis is minimax: stable repeated estimates do not establish proximity to the true entropy because empirical variance does not test bias.Theorem 5 gives a lower bound showing that bias cannot fall below a prescribed δ while n remains below the stated dimension-dependent threshold.Source: §C.4 · p. 15
• For typical β≈0.1, the bias lower bound is at most 0.0057 through d=10^4, but negligible bias requires n at least 20.99^d.With ϵ=0.01, the relevant parameter k⋆ is 3 for d≤11 and 2 for 12≤d≤10^4.Source: §C.4 · pp. 15–16

C.5. Computing the Sample Propagation Estimator

• The Gaussian-mixture entropy needed by the estimator is computed approximately with Monte Carlo integration, under bounded-support or bounded-second-moment assumptions.The method estimates each mixture summand and then averages the resulting log-density terms.Source: §C.5 · p. 16
• The Monte Carlo estimator is unbiased, and its mean-squared-error bounds scale linearly with dimension, with β^2 often dominating experimentally.The bounds are stated for tanh and ReLU networks; the setup assumes finite second moments when support is not bounded.Source: §C.5 · p. 16

D.1. Proof of Theorem 4

• Theorem 4’s proof controls the mutual-information estimation error by bounding conditional entropies uniformly and treating the remaining term as an empirical average.The argument uses independence of injected noise, conditioning-based entropy inequalities, and bounded-support moment control.Source: §D.1 · pp. 16–17

D.2. Proof of Theorem 5

• Theorem 5 rewrites the estimator’s entropy gap as mutual information between propagated samples and a randomly selected noisy sample, then lower-bounds it using a decoding construction.The construction partitions a hypercube into cells, uses maximum-likelihood decoding, and controls the error with Fano’s inequality.Source: §D.2 · pp. 17–18

D.3. Proof of Theorem 6

• The Gaussian-mixture entropy estimator is unbiased, so its MSE reduces to controlling the variance of log g(µ_i+Z).The variance is bounded using the Poincaré inequality for N(0,β^2I_d).Source: §D.3 · p. 18

D.3.1. MSE BOUND FOR BOUNDED SUPPORT √

• Under bounded support, the proof applies a bounded-support inequality and Poincaré control to derive the first MSE bound.The derivation combines the variance estimate with the MSE expansion.Source: §D.3.1 · p. 19

D.3.2. MSE BOUND FOR BOUNDED SECOND MOMENT

• Under a bounded second moment, the proof uses a moment inequality and Poincaré control to derive the second MSE bound.The final step substitutes the variance bound into the MSE expression.Source: §D.3.2 · p. 19