An Information Budget for Label-Noise Memorization, and What Statistical Complexity Tracks: A Study with Multilayer Perceptrons
Authors: Nancy Baygorrea C., Otávio da Fonseca Martins Gomes
Organizations: Laboratory of Mathematical Sciences, State University of Northern Rio de Janeiro · Centre for Mineral Technology, Postgraduate Program in Geosciences, National Museum, Rio de Janeiro
How much label noise has a network memorized, and when? We study a multilayer perceptron and what statistical complexity adds to a description of its predictions. Standard identities decompose predictive information into signal, memorization, and heterogeneity. The memorization term vanishes in the population under the stated noise-independence condition. A Fano-type bound relates it to training accuracies, but the memorized fraction M provides no upper bound. Across 89 MNIST runs with ten pre-specified hypotheses (five supported, four unsupported, one not testable), normalized memorization information rose with M on training samples, remained near zero on held-out samples with fresh noisy labels, and was close to its Fano bound. Unlike raw M, it has a zero population null under the independence condition and chance correction largely removed this difference. When every label was replaced by an incorrect class, test accuracy fell to 0.5%, yet predictions retained information about the true class. Mean statistical complexity peaked before memorization at corruption rates up to 40%. At higher rates, its peak fell within the memorization phase, near the epoch when mean training confidence reached 0.6 and the two could not be separated. As a stopping rule, it matched established rules only at low noise and detected corrupted samples less well than simple label-free scores. An exploratory extension of 280 runs covering Fashion-MNIST, pair-flip noise, and four training variants yielded similar results under uniform noise, but not pair-flip noise, where complexity peaked within two epochs. Conclusions are limited to this architecture, two small datasets, and synthetic noise.
Figures & tables
NAME
RULE FOR A SELECTED SAMPLE
r
RATES USED
UI (uniform over incorrect classes)
y~ drawn uniformly from {1,…,N}∖{y}
ρ
main grid, extension
SU (symmetric uniform)
y~ drawn uniformly from all N classes, so it may equal y
≈ρ(N−1)/N
main grid, ρ=1
PERM (permutation)
the labels of all selected samples are randomly permuted among themselves
≤ρ
main grid, ρ=1
PF (pair-flip)
y~=(y+1)modN
ρ
extension, ρ≤0.4
Table 1: Noise models . For each selected sample of clean class y , the training label y~ is generated as shown; unselected samples keep y~=y . The corrupted fraction r is the fraction of training samples with y~=y .
SYMBOL
MEANING
DEFINED IN
N , n , ny
number of classes; training-set size; training samples of clean class y
Section 2.1
y , y~ , y^
clean, noisy (training) and predicted label; Y , Y~ , Y^ as random variables
Sections 2.1 , 2.4
ρ , r
corruption rate; fraction of training samples with y~=y
Section 2.1
UI , SU , PERM , PF
noise models
Table 1
N , L
corrupted and clean training samples
Section 2.1
pt(⋅∣x) , pi , pˉ , T
predictive distribution at epoch t ; for sample i ; mean prediction; temperature
Sections 2.2 , 2.3
Table 2: Notation . Symbols used throughout the paper and where they are defined.
HYPOTHESIS
DECISION CRITERION
H1
The identities of Propositions 1 and 2 hold numerically
maximum absolute residual <10−8
H2
I(Y^;Y~∣Y) is within its null on test data with fresh noisy labels, and above its null on training data after t10
test: within q95 in ≥90% of scheduled epochs; training: above q99 in ≥95% of scheduled epochs after t10
H3
M(t) and MI(t) agree ( UI , ρ∈{0.2,…,0.8} )
median Spearman correlation over epochs ≥0.9
H4
The peaks of V and of I(Y^;X∣Y~) fall within [t10,t90]
≥80% of runs, for both
H5
Stopping at argmaxt≥1Cˉ loses less test accuracy than the better of the final epoch and the loss plateau
95% CI of the paired difference <0
H6
The per-sample meantCi (first 50 epochs) is non-inferior to the best baseline for detecting corrupted samples
lower bound of the 95% CI of Δ AUROC >−0.02
Table 3: Pre-specified hypotheses and decision criteria . q95 and q99 are quantiles of the within-class permutation null; I(Y^;X∣Y~) is the heterogeneity term computed without clean labels (Proposition 2 ).
ITEM
DEVIATION AND REASON
Memorization times
The pre-specified tq , based on M , is retained, but M has a chance level (Section 5.5 ). Times based on κ and MI are added as exploratory analyses.
H7
H7 was not testable under the protocol because no corrupted run interpolated. It was examined post hoc on SGD runs (Section 5.9 ). The pre-specified criterion was one-sided. We also report the symmetric version.
Trend model
The linear mixed model was replaced by OLS with seed-clustered standard errors because of the boundary fit and supplemented with values for each ρ .
Baselines
Noisy validation and Label Wave were added after the protocol was fixed and are exploratory.
Extension
All 280 extension runs were planned after the main results.
Table 4: Deviations from the pre-specified plan.
SUBJECT
STATISTIC
DECISION
H1
identities
maximum absolute residual =1.2×10−14
pass
H2
null of I(Y^;Y~∣Y)
test: 0.93 of scheduled epochs within q95 ; training: 0.99 above q99 (100 permutations)
Δ AUROC =−0.318 [ −0.381 , −0.255 ] vs. EL2N (60 runs; −0.481 for UI , 0<ρ<1 )
fail
Table 5: Outcomes of the pre-specified hypotheses (definitions and criteria in Table 3 ). Main grid, 80 runs; H1 and H2 evaluated over all 89 runs, including the nine crossed runs. ιsig is normalized by lnN .
ρ
BEST TEST ACCURACY
FINAL TEST ACCURACY
tbest
t10
t50
t90
FINAL TRAIN ACC. (NOISY LABELS)
FINAL M
0
0.986 (0.001)
0.985 (0.001)
109.1
–
–
–
1.000
–
0.2
0.973 (0.001)
0.855 (0.005)
5.6
11.0
18.2
44.2
0.992
0.979
0.4
0.965 (0.002)
0.659 (0.011)
4.9
12.0
20.1
44.5
0.987
0.980
0.6
0.948 (0.002)
0.430 (0.009)
4.1
11.8
20.9
49.8
0.979
0.974
0.8
0.863 (0.007)
0.200 (0.006)
3.1
10.3
23.2
71.8
0.963
0.962
1.0
–
0.005 (0.001)
–
0.5 ∗
20.6
64.3
0.962
0.962
Table 6: Training dynamics under UI noise . Means over 10 seeds and standard deviation in parentheses. Test accuracy is measured against the clean labels. Final M is the memorized fraction at epoch 200. Times are in epochs (Definition 1 ). ∗ At ρ=1 , M starts above 0.1 because of its chance level, so t10 (0 or 1, depending on the seed) is not a memorization time. The corresponding time is 9.6 when based on κ and 15.3 when based on hard MI (Table 9 ). The best test accuracy and tbest exclude epoch 0; at ρ=1 test accuracy stayed below chance from epoch 1 onwards, so these two entries are not reported.
Figure 1: Training accuracy against the noisy labels (left) and test accuracy against the clean labels (right) under UI noise. Mean ± s.d. over 10 seeds. The epoch axis is symmetric-logarithmic.
Figure 2: Memorized fraction M(t) , retained fraction R(t) and test accuracy under UI noise. Shading indicates the memorization phase [t10,t90] , defined using M (medians). Vertical lines indicate the median of the peaks of Cˉ and V and of the best test accuracy.
Figure 3: Left: mean entropy Sˉ (solid) and entropy of the mean prediction S(pˉ) (dashed). Right: the aggregation gap S(pˉ)−Sˉ=I(X;Y^)/lnN (Proposition 1 ). UI noise.
Figure 4: Information budget ( 1 ) on the training set under UI noise, averaged over 10 seeds. The stacked terms sum to S(pˉ) (dashed).
Figure 5: Memorized fraction M(t) and memorization information MI(t) , computed from soft predictions and argmax (hard, temperature-invariant) predictions. UI noise, averaged over 10 seeds.
Figure 6: Budget terms at epoch 200 compared with their complete-memorization values (Proposition 4 , computed from the empirical label channel), for all runs of the main grid and for both soft and hard budgets.
ιsig
ιmem
ρ
REFERENCE
SOFT
HARD
REFERENCE
SOFT
HARD
0.2
0.592
0.590
0.592
0.408
0.382
0.388
0.4
0.326
0.325
0.326
0.674
0.627
0.640
0.6
0.135
0.135
0.136
0.865
0.786
0.809
0.8
0.020
0.019
0.020
0.980
0.846
0.893
1.0
0.046
0.046
0.046
0.954
0.821
0.862
Table 7: Budget terms at epoch 200 compared with their complete-memorization values, computed from the empirical label channel. UI noise, averaged over 10 seeds.
NOISE
ρ
M
Mchance
HARD MI
UI
0.2
0.006
0.005
0.006
UI
0.4
0.008
0.006
0.004
UI
0.6
0.010
0.009
0.004
UI
0.8
0.024
0.021
0.003
UI
1
0.121
0.112
0.003
SU
1
0.105
0.101
0.002
Table 8: Chance level of M after one epoch . Means over 10 seeds. Mchance is the expected value of M if predictions do not depend on which samples were corrupted (Definition 1 ).
t10
t50
NOISE
ρ
M
κ
MI HARD
MI SOFT
M
κ
MI HARD
MI SOFT
UI
0.2
11.0
11.1
12.2
14.0
18.2
18.5
19.9
23.7
UI
0.4
12.0
12.0
13.0
16.0
20.1
20.2
22.6
27.5
UI
0.6
11.8
12.0
14.0
18.0
20.9
21.6
25.0
32.2
UI
0.8
10.3
11.9
15.6
21.9
23.2
25.1
34.0
48.2
UI
1
0.5
9.6
15.3
21.6
20.6
23.1
32.4
45.2
Table 9: Memorization times. Times in epochs, averaged over 10 seeds, exploratory. Times are based on M (pre-specified, after its minimum), chance-corrected κ , and hard and soft MI (without onset correction). Thresholds on MI apply to an information quantity, not a fraction of samples.
Figure 7: M , the hard and soft MI , and the Fano lower bounds of Proposition 5 (dashed, clipped at 0), main grid, mean over 10 seeds. Under SU , PERM and UI at ρ=1 , M starts near its chance level ( ≈0.1 ) while MI starts near 0.
Figure 8: All labels corrupted ( ρ=1 ) under three mechanisms. Left: test accuracy (dotted line: chance). Center: information about the true class in the test predictions, I(Y^;Y)/lnN . Right: Cˉ on the training set. Mean ± standard deviation over 10 seeds.
Figure 9: Components of mean complexity . Decomposition Cˉ=(1−r)CˉL+rCˉN (Observation 1 ) into the contributions of clean and corrupted samples. UI noise, averaged over 10 seeds.
ρ
t∗(Cˉ)
t∗(CˉN)
t∗(CˉL)
SHARE OF Cˉ FROM CORRUPTED
CONF. AT t∗(Cˉ)
Sˉ AT t∗(Cˉ)
0
1.0
–
1.0
0.00
0.958
0.061
0.2
3.7
3.8
3.7
0.20
0.737
0.465
0.4
5.7
5.0
6.2
0.38
0.607
0.629
0.6
19.8
22.0
16.3
0.57
0.583
0.551
0.8
31.8
33.1
28.4
0.80
0.607
0.498
1.0
29.7
29.7
–
1.00
0.591
0.502
Table 10: Peak times of Cˉ and its components, and training confidence at the peak of Cˉ ( UI noise, means over 10 seeds; exploratory). Confidence is the mean of maxkpk on the training set. At ρ=0 there are no corrupted samples, so Cˉ=CˉL . At ρ=1 there are no clean samples, so Cˉ=CˉN .
RULE
ρ=0
ρ=0.2
ρ=0.4
ρ=0.6
ρ=0.8
Final epoch
0.2 [0.1, 0.2]
12.0 [11.6, 12.5]
30.8 [30.0, 31.4]
52.0 [51.4, 52.5]
66.3 [65.7, 66.9]
Loss plateau
0.8 [0.6, 1.0]
12.1 [11.5, 12.6]
28.9 [28.2, 29.6]
50.2 [49.0, 51.4]
65.1 [64.7, 65.6]
argmaxCˉ (pre-specified)
1.9 [1.7, 2.0]
0.2 [0.1, 0.3]
0.4 [0.2, 0.6]
31.0 [28.6, 33.2]
60.0 [59.4, 60.7]
argmaxV
1.9 [1.7, 2.0]
7.8 [7.3, 8.3]
25.1 [24.5, 25.8]
46.1 [45.3, 47.1]
64.7 [64.1, 65.5]
Noisy validation, 5000
0.2 [0.1, 0.3]
0.4 [0.2, 0.6]
0.2 [0.1, 0.4]
0.2 [0.0, 0.4]
0.8 [0.2, 1.5]
Noisy validation, 1000
0.3 [0.3, 0.3]
0.4 [0.3, 0.4]
0.3 [0.2, 0.4]
0.4 [0.3, 0.6]
2.0 [1.6, 2.6]
Table 11: Early-stopping evaluation . Loss in clean test accuracy relative to the best epoch, evaluated on a held-out half of the test set under UI noise. Values are percentage points, reported as means and 95% bootstrap CI over 10 seeds. All nine Label Wave configurations are reported in Table 15 .
Figure 10: Early-stopping benchmark . Top: loss in test accuracy for each stopping rule versus the corruption rate (mean and 95% CI). Bottom: median stopping epoch of each rule, shown on the mean test-accuracy curve for the evaluation half. Shading indicates the memorization phase defined using M .
W=10
W=25
W=50
SCORE
ρ=0.2
0.4
0.6
0.8
ρ=0.2
0.4
0.6
0.8
ρ=0.2
0.4
0.6
0.8
entropy
0.693
0.662
0.642
0.598
0.838
0.751
0.679
0.588
0.831
0.730
0.656
0.571
largest probability
0.690
0.659
0.642
0.605
0.883
0.805
0.726
0.612
0.889
0.786
0.697
0.587
top-2 margin
0.687
0.656
0.639
0.604
0.902
0.830
0.746
0.620
0.913
0.808
0.708
0.588
temporal variance
0.619
0.518
0.445
0.438
0.905
0.799
0.630
0.480
0.969
0.936
0.875
0.589
prediction changes
0.554
0.549
0.556
0.572
0.884
0.831
0.775
0.657
0.928
0.876
0.806
0.659
Table 12: Label-free detection of corrupted samples (exploratory): AUROC with the direction fixed in advance, for scores averaged over the first W epochs ( UI noise; means over 10 seeds, standard deviations at most 0.014). The last row uses the noisy label and is shown for reference. With W=50 the logits are available at 40 epochs (every second epoch after epoch 30).
Figure 11: Detection of corrupted samples . AUROC, with the direction fixed in advance, for identifying corrupted training samples from features computed over the first 50 epochs (mean ± s.d. over 10 seeds). Geometric features (prefix geo : mean and standard deviation of Ci over epochs, mean of Si and Fisher-Rao path length) depend only on pi . The baselines use the noisy label: area under the margin ( Pleiss et al., 2020 ) , forgetting events ( Toneva et al., 2019 ) , mean loss, EL2N ( Paul et al., 2021 ) and dataset-cartography confidence and variability ( Swayamdipta et al., 2020 ) .
SPEARMAN rs
CONDITION
ρ
INTERP.
t∗(Cˉ)
tbest
FINAL
argmaxCˉ
NOISY. VAL. 1000
LABEL WAVE k1p3
MNIST, UI , Adam (protocol)
0.2-0.8
0.00
0.93
−0.67
40.3
22.9
0.8
0.4
MNIST, UI , Adam cosine
0.2-0.8
0.75
0.93
−0.58
38.3
23.0
0.7
0.3
MNIST, UI , Adam w.d. 10−4
0.2-0.8
0.03
0.94
−0.68
39.5
23.6
0.8
0.3
Table 13: Robustness summary . Exploratory analyses, with 10 seeds per rate. Spearman ( rs ): correlation of the event time with ρ over 0<ρ<1 . Early stopping values are mean losses in clean test accuracy, in percentage points, over 0<ρ<1 , evaluated on the evaluation half of the test set. Interpolation is the fraction of corrupted runs reaching 99.9% training accuracy against noisy labels.
Figure 12: End-point robustness . Budget terms at the final epoch compared with their complete-memorization values across all conditions (circles: ιsig ; triangles: ιmem , soft). The ιmem points on the diagonal at high noise correspond to runs that reached interpolation (SGD and cosine schedule).
Figure 13: Timing across conditions . Peak of Cˉ ( tCˉ∗ ), peak of V ( tV∗ ) and best test epoch versus ρ , by condition (medians over seeds). Shading indicates the memorization phase defined using M under the pre-specified definition. Under pair-flip noise, its lower boundary is affected by the chance level of M (Section 5.5 ). For SGD at ρ=0.4 the peak of V is bimodal across runs (epoch 1 in six runs, 53-60 in four).
Figure 14: Early stopping across conditions . Loss in clean test accuracy, in percentage points, for four stopping rules versus ρ , by condition (means over seeds).
Figure 15: Discrimination after interpolation . Run-level AUROC of training-set Cˉ for distinguishing runs with ρ=0 from runs with ρ>0 , by epoch. Left: SGD, for which all runs reached interpolation and clean runs did so at a median epoch of 18 and noisy runs at median epochs of 113-122. Right: the Adam protocol, under which no corrupted run reached interpolation. The dotted line indicates the pre-specified threshold of 0.6.
RQ
DECISION
INTERPRETATION
H1
–
pass
Implementation check of exact identities.
H2
RQ1
pass
The memorization term behaved as expected on held-out and training data; on held-out data the null holds by construction, so this checks calibration.
H3
RQ1
pass
MI and M agreed in ordering over epochs, a lenient criterion; their timings differed (Table 9 ).
H4
RQ2
fail
The peak of V fell within a wide memorization window; that of I(Y^;X∣Y~) did not single out the phase.
H5
RQ2
pass
The peak of Cˉ beat lenient baselines, but not established rules at ρ≥0.6 .
H6
RQ2
fail
Per-sample complexity was weaker than label-based and label-free scores.
Table 14: Hypotheses, outcomes and interpretation . Decisions follow the pre-specified criteria (Table 3 ).
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 16: Centroid trajectories in the entropy-disequilibrium plane . Trajectories of (Sˉ(t),Dˉ2(t)) under UI noise, averaged over seeds, from epoch 0 (open circle) to epoch 200 (square). The shaded area is the region attainable by single distributions for N=10 , bounded above by γsup (solid) and below by γinf (dotted). Crosses mark, for each ρ , the point of the Bayes posterior of the noisy label under UI noise, which assigns 1−ρ to the clean label and ρ/9 to each other class. A network predicting this posterior for every sample would reach it. Dashed lines are level sets of S⋅D2 ( 0.05 , 0.1 , 0.2 ), not of Cˉ .
(k,p)
ρ=0
ρ=0.2
ρ=0.4
ρ=0.6
ρ=0.8
ρ=1
(1, 3)
0.53
0.22
0.14
0.15
1.05
0.34
(1, 5)
0.41
0.22
0.14
0.15
1.05
0.28
(1, 10)
0.20
0.22
0.14
0.15
1.05
0.20
(3, 3)
0.58
0.41
0.33
0.78
3.03
0.30
(3, 5)
0.54
0.41
0.33
0.78
3.03
0.28
(3, 10)
0.34
0.41
0.33
0.78
3.03
0.22
Appendix
Table 15: Results for the nine Label Wave configurations . Loss in clean test accuracy relative to the best epoch, in percentage points, averaged over 10 seeds and evaluated on the evaluation half of the test set under UI noise. The rows (1,3) and (3,5) are those reported, rounded, in Table 11 . For 0<ρ<1 , configurations with the same window k gave identical losses for every patience p ; the patience changed the result only at ρ=0 and ρ=1 .
Figure 17: Components of statistical complexity . Sˉ , Dˉ2 and Cˉ on the training set under UI noise, mean ± standard deviation over 10 seeds (Section 5.7 ).
Figure 18: Within-group variance . Left: within-group variance V , with groups defined by the noisy label (logarithmic scale). Right: V normalized by the bound of Observation 1 , with Cmax=0.2519 for D2 (Appendix A ). UI noise, mean ± standard deviation over 10 seeds (Section 5.7 ).
Figure 19: Hard information budget . Information budget computed from argmax predictions under UI noise, averaged over 10 seeds (Section 5.4 ). Because argmax predictions are a deterministic function of the input, the mean conditional entropy is zero, the heterogeneity term equals H(Y^∣Y,Y~)/lnN , and the three terms add up to the normalized entropy of the predicted labels. These quantities do not change under temperature scaling.
Figure 20: Label Wave signal and memorized fraction . Top: fraction of training samples whose predicted class changes between consecutive epochs (the Label Wave signal; logarithmic scale). Bottom: memorized fraction M(t) . Curves are means over 10 seeds under UI noise. Vertical lines mark the median over seeds of the Label Wave stopping epoch ( k=3,p=5 ; dashed) and of the best epoch on the evaluation half of the test set (dotted) (Section 5.8 ).
Figure 21: Noisy-validation size and rule selection ( UI noise). Left: loss in clean test accuracy of the noisy-validation rule as a function of the validation-set size (solid lines and markers), and of stopping at the peak of Cˉ (dashed line of the same color), for each ρ ; means over 10 seeds. Right: loss of the rule selected on five seeds and evaluated on the other five, compared with the peak of Cˉ and with noisy validation; bars are means with 95% bootstrap intervals, and the numbers above the bars are the means. In this panel, selection and evaluation use the same test images; the selection on separate halves of the test set, reported in Section 5.8 , identified the same rule. Both panels use a symmetric-logarithmic vertical axis.
Neural network memorization studies how models store and retrieve training examples, how this differs from generalization, and how memory can be measured or controlled.
Highly over-parameterized models can simultaneously memorize noisy labels and generalize well, yet how these behaviors coexist remains poorly understood. In this work, we investigate the underlying mechanisms of this coexistence using modular arithmetic tasks under heavy label noise. Through extensive experiments on two-layer neural networks, we find that larger models tend to generalize better under appropriate optimization and model configurations, while noisy labels are memorized faster than clean data. Over-parameterized models internally form a generalization structure, but its expression in the output is suppressed by the need to fit noisy labels. Remarkably, even with 80% label noise, near-perfect test accuracy can be achieved by extracting this internal structure using frequency-based methods. We further propose a task-agnostic method to partition networks into generalization and memorization components. Although this subnetwork improves generalization, it is limited compared with frequency-based extraction, indicating that the generalization structure is distributed across neurons and motivating the development of new tools to retrieve generalizable knowledge from over-parameterized networks.
Linyu Liu, Pinyan Lu
1Taylor Lab, Huawei Technologies Co., Ltd. · 2Key Laboratory of Interdisciplinary Research of Computation and Economics, Shanghai University of Finance and Economics
We introduce a straightforward yet effective method to empirically study memorization in deep neural networks for classification tasks. Our approach augments each training sample with auxiliary random labels, which are then predicted by a random label prediction head (RLP-head). RLP-heads can be attached at arbitrary depths of a network, predicting random labels from the corresponding intermediate representation and thereby enabling analysis of how memorization capacity evolves across layers. By interpreting the RLP-head performance as an empirical estimate of Rademacher complexity, we obtain a direct measure of both sample-level memorization and model capacity. We leverage this random label accuracy metric to analyze generalization and overfitting in different models and datasets. Building on this approach, we further propose a novel regularization technique based on the output of the RLP-head, which demonstrably reduces memorization. Interestingly, our experiments reveal that reducing memorization can either improve or impair generalization, depending on the dataset and training setup. These findings challenge the traditional assumption that overfitting is equivalent to memorization and suggest new hypotheses to reconcile these seemingly contradictory results. The source code is available at https://github.com/MarlonBecker/RandomLabelHeads
Marlon Becker, Jonas Konrad, Luis Garcia Rodriguez +1
A fruitful approach towards understanding generalization in machine learning involves characterizing how models separate generalizable structure from idiosyncratic or even corrupted features of training data. Fisher information has emerged as a diagnostic tool for this separation, leveraging loss geometry to distinguish generalization from memorization. Here we investigate this connection in vision models trained on CIFAR-10 with controlled noise that can be learned only through memorization. We show that projecting model weights onto the top Fisher eigenvectors decouples general task performance from noisy sample memorization, and we investigate how this decoupling presents differentially across layers. We show that the same phenomena are present in small transformers trained to solve a synthetic in-context learning task, and are reproduced in closed form in a simple, analytically solvable linear model. Our results reveal that Fisher information captures aspects of the geometric structure underlying how neural networks allocate capacity between generalization and memorization.
Daniel T. Bernstein, Chase Goddard, Luca Di Carlo +2
Princeton University1; Lewis-Sigler Institute · Princeton University2; Joseph Henry Laboratories · University of Sydney3; School of Physics +1