REVIEW 5 major objections 6 minor 6 cited by
A Closer Look at Multimodal Representation Collapse
T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read Modality collapse, the paper argues, happens when polysemantic fusion-head neurons entangle one modality's noisy features with another's predictive ones; freeing the resulting rank bottleneck—by distillation or explicit basis…
desk verdict A useful empirical recipe (EBR) and a plausible mechanism, but the theoretical core has a vacuous definition and asymptotically vacuous bounds; treat the theorems as conjectures. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
Polysemantic neurons—fusion-head neurons that encode features from more than one modality—are the mechanism's primary object. Definition 1 introduces 'conjugate features': a pair $z$, $z^*$ in the same modality, at least one predictive, whose mutual information with the label cancels when they are encoded together, i.e. $I(z;y)+I(z^*;y)=0$ and $I(zz^*;y)=0$; Theorem 1 claims these cancelling pairs are exactly what makes the entangled modality's marginal loss contribution vanish. Lemma 1 quantifies the collision rate: the fraction of cross-modal polysemantic neurons grows quadratically in the number of modalities. The rank bottleneck is the second load-bearing piece: Lemma 2 states that SGD gradient updates converge in rank toward the average gradient outer product (AGOP), and Theorem 2 bounds the distance between any weight subspace $w$ and the AGOP by $\gamma(w)^{-1/n}$, tighter for more polysemantic subspaces—so SGD preferentially settles into polysemantic neurons. Theorem 3 says that once the encoders are aligned by knowledge distillation, that bound loosens to $\kappa^{-1/n}$, polysemantic subspaces shrink, and monosemantic neurons become viable. Explicit Basis Reallocation carries the practical argument: an encoder-decoder pair per modality plus a modality discriminator, trained adversarially, so each modality's predictive features occupy the full output basis and noisy features get dedicated dimensions to be denoised along.
What would settle it
In a model that exhibits collapse, compute empirical mutual information between each fusion-head feature and the label: the theory requires a pair $(z, z^*)$ with $I(z;y)+I(z^*;y)=0$ while at least one is predictive, which non-negativity of mutual information makes impossible unless both are zero—so finding no such pair would empty Theorem 1 of its subject. Alternatively, train the identical fusion architecture with spectral or orthogonal regularization that provably removes the rank bottleneck: if collapse still occurs, rank starvation is not the driver the paper claims.
Extended reading notes
Core claim
On the paper's own terms: modality collapse is not primarily a symptom of conflicting gradients or data-distribution mismatch, but of an unintended cross-modal entanglement inside the fusion operator. Because neurons are polysemantic—each encoding features from more than one modality—and because the probability that a fusion-head neuron simultaneously encodes a noisy feature of one modality and a predictive feature of another grows quadratically with the number of modalities, the predictive features of the weaker modality get numerically cancelled by the noisy features entangled with them. Gradient updates under SGD are rank-constrained, converging toward the average gradient outer product (AGOP), which biases the fusion head toward exactly such low-rank polysemantic parameterizations. Cross-modal knowledge distillation implicitly loosens this bottleneck: as student representations are pulled toward the teacher's, the fusion head's polysemantic subspaces shrink, freeing dimensions along which the noisy features can be denoised independently. The paper's algorithm, Explicit Basis Reallocation (EBR), turns this implicit process into an explicit adversarial objective, and the resulting per-modality basis mapping supports test-time substitution of missing modalities.
Load-bearing premise
The central mechanism presupposes 'conjugate' feature pairs that are predictive yet cancel each other out in information-theoretic terms, and the rank-bottleneck bounds presuppose that every feature carries equal unique label information; if such pairs cannot exist or the equal-information condition fails, the collapse mechanism has no subject matter.
Editorial extensions
If this is right
- Collapse should become more likely as the number of modalities grows, since cross-modal polysemantic collisions scale quadratically with the number of modalities; models with many input channels will need explicit rank management.
- Distilling the surviving modality into the collapsed one, sequenced from weakest to strongest, should raise the effective rank of the student encoder and restore the collapsed modality's contribution; the paper reports this pattern on MIMIC-IV and avMNIST.
- Because EBR modifies only the unimodal encoders, every fusion strategy should improve out-of-the-box when its backbones are swapped for KD or EBR versions; the paper reports such gains on MUSE, SMIL, ShaSpec, Grape, and M3Care.
- The feature-to-basis mapping learned by EBR doubles as a substitutability ranking: when a modality is missing at test time, another modality's encoding can be projected through the decoder to stand in for it, which the paper validates on MIMIC-IV missingness rates up to 0.7.
Reading between the lines
- Any intervention that raises the effective rank of gradient updates in the fusion head—wider heads, spectral or orthogonal regularization, gradient whitening—should also suppress collapse, and the phase-transition point observed near $\beta=4$ in the paper's rank experiments should shift accordingly; this is a testable prediction the paper does not run.
- The formal definition of conjugate features appears problematic: since mutual information is non-negative, $I(z;y)+I(z^*;y)=0$ forces both terms to zero, contradicting the requirement that at least one feature be predictive; the qualitative story of noisy interference may still be correct, but it needs a different formalization, for instance in terms of cancelled gradients rather than cancelled mu
- The appendix's variance inflation factor measurements suggest a cheap monitoring proxy: if basis reallocation works by de-correlating features across modalities, then collapse severity should track measurable multicollinearity of the fused representation, and VIF or effective rank could serve as early-warning metrics during training.
- The substitution intuition implies that the encoders share a latent factor space up to equivariance; a natural stress test the authors do not run is whether EBR's missing-modality ranking survives distribution shift between training and test missingness patterns.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper aims to provide a mechanistic theory of modality collapse in multimodal fusion, arguing that noisy features from one modality become entangled with predictive features from another through shared polysemantic neurons in the fusion head, thereby suppressing the predictive contribution of the former modality. It formalizes this via conjugate features and an interference theorem (Theorem 1), links it to a low-rank simplicity bias through gradient-rank limits and a polysemantic bottleneck bound (Lemma 2, Theorem 2), and claims that cross-modal knowledge distillation frees rank bottlenecks (Theorem 3). On this basis it proposes Explicit Basis Reallocation (EBR), an adversarial encoder-level method, and reports improved missing-modality results on MIMIC-IV and avMNIST over several baselines.
Significance. If the theory were sound, the paper would offer a novel and practically relevant account of modality collapse, with a simple plug-in remedy (EBR) that is fusion-agnostic. The empirical portion is substantial: experiments span two datasets, multiple fusion backbones, missing-modality rates, noise perturbations, distillation-order ablations, and statistical significance tests; the rank and representation-similarity measurements are a useful contribution independently of the theory. However, the theoretical core is load-bearing for the paper's central claim, and the concerns raised by the stress-test analysis are confirmed by the manuscript text: Definition 1 is vacuous as written, the proof of Theorem 1 assumes the cancellation it purports to derive, and Theorems 2 and 3 contain bounds whose right-hand sides do not behave as claimed and whose key quantities are not well defined. The empirical results may stand on their own, but they cannot rescue the stated theoretical contribution.
major comments (5)
- [Section 3.1 (Definition 1)] The conjugate-feature definition is internally inconsistent and makes Theorem 1's premise empty. Since mutual information is nonnegative, the condition I(z;y)+I(z*;y)=0 forces I(z;y)=I(z*;y)=0, so neither z nor z* carries predictive information about y. This contradicts the requirement that at least one of z,z* has predictive value. The additional condition I(zz*;y)=0 is then redundant. Consequently, the object z_y ('predictive conjugate features') used in Theorem 1 does not exist under the stated definition, and the paper's central interference mechanism is not established. The definition would need to be reformulated (e.g., in terms of cancellation of pre-activations rather than mutual-information equalities) before Theorem 1 can have subject matter.
- [Section 3.1 (Theorem 1 proof)] The proof asserts, rather than derives, the geometric cancellation that drives the result. It states that because z_y and z_epsilon are conjugate, z_y activates in the opposite direction to z_epsilon along the polysemantic subspace w_p with large equal magnitudes, yielding w_p·z_y + w_p·z_epsilon = 0. Neither the sign opposition nor the equal-magnitude condition follows from Definition 1, which only imposes mutual-information equalities, nor from Lemma 3, which bounds a ratio of sums of dot products over weight subspaces. The limit in the theorem is therefore a restatement of the assumed cancellation, not a consequence of the preceding lemmas.
- [Section 3.2 (Theorem 2)] The polysemantic bottleneck bound has several load-bearing defects. First, the left-hand side is not well defined: w is a subspace of the weight matrix, while the AGOP is a matrix, and no norm or embedding that places a subspace and a matrix in a common space is specified. Second, for any fixed gamma(w)>0, gamma(w)^{-1/n} tends to 1 as n grows, so the claimed bound becomes looser, not tighter, with training; it cannot establish that W converges to high-polysemanticity subspaces. Third, the chain K(1-2mu*lambda)^{nl} <= gamma(w)^{-1/n} in Eq. (4) compares a geometrically decaying quantity with a quantity that increases to 1, and no justification is given for the inequality. The interpretation in Section 4.2 that the rank drop 'is a consequence of the upper-bound in Theorem 2' is therefore not supported.
- [Section 3.3 (Theorem 3)] Theorem 3's conclusion is built into its construction rather than derived. In the proof, the authors assume |w∩X|→k_w and then define kappa=k_w/dim(w), which is exactly the degree of polysemanticity gamma(w) of the subspace. The claim that alignment shrinks polysemantic subspaces is thus an assumption about the boundedness of |w∩X|, not a consequence of knowledge-distillation dynamics. The theorem also inherits the vacuous-limit issue of Theorem 2, since kappa^{-1/n}->1 as n→∞. In addition, the 'weighted general case' of the Depth-Rank Duality is introduced in the proof without a formal statement or derivation, and the explicit form of alpha=h_phi(d(x~_i,x~_j)) and its monotonicity alpha1>alpha2>... are assumed without evidence.
- [Section 3.2 (Lemma 2)] Lemma 2 is not proved in the manuscript; its proof consists of two citations, one for rank decrease under SGD and one for convergence to the average gradient outer product. The statement rank(∇_l L_n) proportional to rank(Σ_x ∇φ_l(x)∇φ_l(x)^T) is not derived for the multimodal fusion architecture, and no rate or uniformity over n is given. Since Theorem 2 and the empirical rank-bottleneck interpretation in Section 4.2 depend on this lemma, the rank-bottleneck mechanism should be treated as conjectural rather than established.
minor comments (6)
- [Section 3.4] The phrase 'compete output basis' should read 'complete output basis'.
- [Table 1] The MUSE AUC-PRC entry is malformed ('0.39.87 ± 0.05' should presumably be '0.3987 ± 0.05').
- [Definition 1] The notation I(zz*;y) is ambiguous; it should be written as I((z,z*);y) with the joint variable made explicit.
- [Definition 2 and Theorem 2] The quantity |w∩X| is not formally defined: w is a weight subspace and X is an input distribution, and no formal intersection operation is specified.
- [Section 5] The limitation that Theorems 2 and 3 assume equal conditional cross-entropy reduction across features is acknowledged only in the conclusion; it should be stated as a hypothesis before the theorems are introduced.
- [Lemma 3] The lemma statement contains the typo 'predicitvity', and the proof invokes the Johnson–Lindenstrauss lemma through a reference to Elhage et al. (2022) without stating the specific property being used.
Circularity Check
Definition 1 empties Theorem 1 (nonnegative MI forces conjugate features to have zero label information), and Theorem 3's κ-bound is Theorem 2's polysemanticity γ relabeled; both rest on the authors' own Depth-Rank Duality.
-
self definitional
[Section 3.1, Definition 1; used in Theorem 1]
"Definition 1 (Conjugate Features). A conjugate feature z is one that coexists, in a given modality, with another feature z∗ such that at least one of them has some predictive value, but they can semantically cancel each other out when considered in conjunction, i.e., I(z; y) + I(z∗; y) = 0; I(zz∗; y) = 0"
Mutual information is non-negative, so I(z; y) + I(z*; y) = 0 forces I(z; y) = I(z*; y) = 0. Thus the 'at least one has predictive value' requirement is unsatisfiable, and every 'predictive conjugate feature' zy has zero marginal and joint MI with y. A zero-MI feature contributes nothing to loss reduction, so the Theorem 1 limit Σ ∂L/∂wp = 0 holds by definition of zy, regardless of p(wp). The proof's additional claim that zy 'activates in the opposite direction' of zϵ with large magnitude is an extra geometric assumption, not a consequence of the MI equalities. The central interference mechanism is therefore a tautology over an empty set rather than a derived prediction.
-
self definitional
[Appendix B.2, proof of Theorem 3 (Dynamic Convergence Bound)]
"Since both |wϵ| and |X| are constants, kw is also a constant. It thus follows that, under a dynamic input space approaching a bounded neighborhood ϵ, the number of features encoded in any polysemantic subspace also gets bounded by kw as n → ∞. So, the RHS in Theorem 2 can be rewritten as: lim n→∞ γ(w)^−1/n = (|w ∩ X| / dim(w))^−1/n = (kw / dim(w))^−1/n = κ^−1/n"
Definition 2 defines γ(w) = |w ∩ X| / dim(w). The proof of Theorem 3 defines κ = kw / dim(w), where kw is exactly the feature count per subspace after alignment. So κ^−1/n is the same object as γ(w)^−1/n with the numerator replaced by the assumed limit kw. The claim that alignment 'bounds' the number of features per neuron is precisely the decrease in polysemanticity that Theorem 3 purports to establish; substituting it into Theorem 2's bound yields the theorem statement. The conclusion that distillation frees rank bottlenecks and increases monosemanticity is therefore the limit assumption d→ϵ restated as a bound, not derived from an independent mechanism.
1 more flagged steps
-
self citation load bearing
[Appendix B.2, proof of Theorem 3; Theorem 4 cited in Appendix B.1]
"A weighted general case of the Depth-Rank Duality result by (Sreelatha et al., 2024) follows as a consequence of cross-modal interferences, which can be written as: min_{φ,f} L(φ(X), Y) + α min_φ Σ_l rank(φ)"
The Depth-Rank Duality is imported from Sreelatha et al. (2024), a prior paper coauthored by two of the present authors (Chaudhuri and Dutta). That theorem supplies the rank-minimization objective on which both Theorem 2 ('as formalized in Theorem 4 by Sreelatha et al. (2024)') and Theorem 3 rest. The 'weighted general case' is declared to follow from the self-cited result rather than proved here; no external, machine-checked, or independently reproduced statement is provided. Since the rank bottleneck is the paper's explanation for polysemanticity and collapse, the load-bearing premise is a self-citation chain rather than a first-principles derivation.
full rationale
The paper's own equations make the central theoretical mechanism circular. Definition 1 defines conjugate features by the equality I(z; y) + I(z*; y) = 0; because mutual information is nonnegative, this forces both marginal MIs to zero, contradicting the requirement that at least one has predictive value and making the 'predictive conjugate features' zy of Theorem 1 an empty class. The theorem's vanishing-gradient limit then follows by definition, not from polysemantic collisions. The proof adds an unproved geometric cancellation assumption (opposite activation along wp). Theorem 3 is similarly self-definitional: its κ is defined as kw/dim(w), which is exactly the polysemanticity degree γ(w) after the alignment limit, so the 'dynamic convergence bound' is Theorem 2's bound restated with the desired conclusion—that alignment reduces feature count per neuron—already assumed. Both theorems lean on Depth-Rank Duality from Sreelatha et al. (2024), prior work coauthored by two of the present authors; the weighted generalization used in Theorem 3 is asserted, not derived. The experimental section and EBR algorithm are independent contributions and could support an empirical claim, but the paper's abstract-level theoretical explanation reduces to these definitional and self-cited steps. Score is therefore 8: the central derivation is forced by definition and by a self-citation chain, while the experiments and EBR algorithm prevent the maximum score.
Assumptions & free parameters
assumptions (4)
- domain assumption The number of task-relevant features in X is greater than the number of neurons in any layer of the fusion head (polysemanticity presupposition).
- ad hoc to paper Conjugate features exist as defined in Definition 1, i.e., a pair z, z* with I(z;y)+I(z*;y)=0 and I(zz*;y)=0 while at least one has predictive value.
- ad hoc to paper A weighted general case of the Depth-Rank Duality (Sreelatha et al., 2024) holds and provides a bound on rank regularization.
- domain assumption SGD weight matrices converge to a subspace proportional to the average gradient outer product (AGOP).
Cite this review
Pith. "Pith review of A Closer Look at Multimodal Representation Collapse." pith.science (2026). https://pith.science/paper/AX4G7K4T
@misc{pith2026250522483,
author = {Pith},
title = {Pith review of: A Closer Look at Multimodal Representation Collapse},
year = {2026},
howpublished = {\url{https://pith.science/paper/AX4G7K4T}},
note = {Machine review of arXiv:2505.22483}
}
read the original abstract
We aim to develop a fundamental understanding of modality collapse, a recently observed empirical phenomenon wherein models trained for multimodal fusion tend to rely only on a subset of the modalities, ignoring the rest. We show that modality collapse happens when noisy features from one modality are entangled, via a shared set of neurons in the fusion head, with predictive features from another, effectively masking out positive contributions from the predictive features of the former modality and leading to its collapse. We further prove that cross-modal knowledge distillation implicitly disentangles such representations by freeing up rank bottlenecks in the student encoder, denoising the fusion-head outputs without negatively impacting the predictive features from either modality. Based on the above findings, we propose an algorithm that prevents modality collapse through explicit basis reallocation, with applications in dealing with missing modalities. Extensive experiments on multiple multimodal benchmarks validate our theoretical claims. Project page: https://abhrac.github.io/mmcollapse/.
Figures
Figures from the paper (6 more)
Forward citations
Cited by 6 Pith papers
-
Feeling the Unexpected: ResTacVLA for Contact-Rich Manipulation via Residual Tactile Representation
Residual tactile representations plus surprise-aware gating let VLA policies master contact-rich robot tasks that vision-only models fail.
-
General Incomplete Multimodal Learning via Dynamic Quality Perception
A unified multimodal learning framework models modality degradation as a continuous variable and uses a noise-aware quality estimator to adaptively weight fused representations under both intra- and inter-modality mis...
-
Escaping the Procrustean Bed: Groupwise Orthogonal Connectors for Audio-Language Models
ORCA splits Q-Former queries into orthogonally constrained groups, reversing directional collapse and speaker-indistinguishability in audio-LLM connectors and gaining 26.4 points on SAKURA multi-hop reasoning.
-
Feature-level Interaction Explanations in Multimodal Transformers
FL-I2MoE separates unique, synergistic, and redundant cross-modal evidence at the token/patch level and uses SII and redundancy-gap scores to rank pairs whose removal degrades performance more than random masking.
-
Diverse via bounded Agreement: Geometric Regularization for Multimodal Fusion
A regularization method enforces diverse intra-modal embeddings and bounded inter-modal drift to improve both multimodal fusion and unimodal robustness.
-
Data Curation Matters: Model Collapse and Spurious Shift Performance Prediction from Training on Uncurated Text Embeddings
Training on LLM text embeddings can cause tabular classifiers to collapse to single-class predictions, which spuriously inflates Accuracy-on-the-Line correlations.
Reference graph
Works this paper leans on
-
[1]
Lemma 3 (Entanglement by Feature Type)
(dim fmin)2 mX i=1 dim fi !2 This completes the proof of the lemma. Lemma 3 (Entanglement by Feature Type). If the latent factors underlying the modalities are sufficiently complementary to 12 A Closer Look at Multimodal Representation Collapse each other in terms of predicitvity of the label y, i.e., for any pair of modalities i and j, X p X q zp i · zq ...
work page 2022
-
[3]
Galanti, T., Siegel, Z. S., Gupte, A., and Poggio, T. A. SGD and weight decay secretly minimize the rank of your neural network. In NeurIPS 2024 Workshop on Mathematics of Modern Machine Learning,
work page 2024
-
[4]
Therefore, the combination with a steep entry must lead to a lower rank solution. When the equality is not met across all features, the low-rank / steepness condition is trivially satisfied by the existence of a lower-dimensional subspace of zis that has a lower conditional mutual information I(x; y|zi), and deriving the upper-bound on the rank in terms o...
work page 2022
-
[6]
MIMIC-IV: Comparison of the substitutability performance of EBR with baselines. Experiments performed over a subset of 4 (2 strong and 2 weak) input modalities, based on which, the average of 15 possible missingness patterns with standard deviation are reported. trained on specifically on the subset of modalities that do not go missing, i.e., its performa...
work page 2024
-
[8]
P-values of the Wilcoxon rank test with Holm–Bonferroni correction on the Table 1 results between EBR and other baselines. C.7. Polysemanticity Considering the results in Figure 5 (a) and (c), Figure 7, and Section 4.3, since there is no external source of noise in the fusion head, and encouraging monosemanticity through basis reallocation has a denoising...
work page 2022
-
[9]
We adopt the experimental setup of Wang et al. (2023) for avMNIST. Implementation Details: The two hidden layers of ψ have output dimensionalities 512 and 256 respectively. The hidden layers of h have output dimensionalities 1024 and 512 respectively, whereas that ofh−1 is 512 and
work page 2023
-
[10]
This completes the proof of the lemma
With the elimination of noisy features, such entanglements following from superposition become less likely among predictive features across modalities, since they would normally require dedicated dimensions with strong orthogonality,i.e., monosemantic neurons. This completes the proof of the lemma. Theorem 1 (Interference). As the number of cross-modal po...
work page 2022
-
[11]
and that this minimization is more pronounced as we go deeper into the neural network (Huh et al., 2023), as formalized in Theorem 4 by Sreelatha et al. (2024). In other words, the deeper we go into a network, the more likely it is for the representations to be of a lower rank (Huh et al., 2023; Sreelatha et al., 2024), and that this rank decreases with e...
work page 2024
Show all 21 references
-
[12]
of the corresponding neurons by SGD, which follows from the Johnson–Lindenstrauss lemma (Elhage et al., 2022). It happens because SGD, by default, performs the aforementioned implicit weighted rank-regularization aside from ERM, with the highest possible value of α such that i...
2022
-
[17]
avMNIST: Comparison with SOTA on dealing with the missing audio modality across different missingness rates at test time, following the baseline setup of Wang et al. (2023). C.4. Baselines for Substitutability We design the following baselines for comparison against our EBR-ba...
2023
-
[20]
As we can see, trends similar to that of our original setting reported in the main manuscript, in the semantic loss gap between the Multimodal Prefix and the Unimodal Baseline, play out when we perform a contrastive objective based fusion as reported in Poklukar et al. (2022)....
2022
-
[21]
as our candidate generative model for analyzing rank bottlenecks. Since the objective of generative modeling is somewhat different from the downstream application that we experimented with, to analyze MMV AE, we performed the experiment on their proposed MNIST-SVHN dataset, wh...
2022
-
[1995]
Chaudhuri, A., Mancini, M., Chen, Y ., Akata, Z., and Dutta, A
doi: 10.1162/neco.1995.7.1.108. Chaudhuri, A., Mancini, M., Chen, Y ., Akata, Z., and Dutta, A. Cross-modal fusion distillation for fine-grained sketch- based image retrieval. In BMVC,
1995 doi
-
[1998]
and the Free Spoken Digits Dataset (Jackson et al., 2018), where the task is to predict the labels of the 18 A Closer Look at Multimodal Representation Collapse input digits from 0 to
2018
-
[2017]
S., Benton, J., and Shlegeris, B
Scherlis, A., Sachan, K., Jermyn, A. S., Benton, J., and Shlegeris, B. Polysemanticity and capacity in neural networks. ArXiv, abs/2210.01892,
-
[2019]
Intuition behind the Dynamic Convergence Bound: Every modality consists of both noisy and predictive features
subspace. Intuition behind the Dynamic Convergence Bound: Every modality consists of both noisy and predictive features. If fusion collapses to a specific modality (target), it means that the modality contains more predictive information and less noise than the rest. Knowledge...
2024
-
[2020]
and autoencoders (Baldi, 2012), which the authors found to be similarly competitive. Although there have been some works that explored the applicability of knowledge distillation to dealing with missing modalities, their purposes have been scoped to addressing issues such as c...
2012
-
[2021]
based approaches attempted to operate without this assumption. Other approaches such as those that facilitate direct interaction among modality-specific raw inputs (Kim et al., 2021; Lee et al., 2023), and ones based on self-supervised domain adaptation (Shen & Gao, 2019), als...
2022
-
[2022]
Invariant risk minimization
Arjovsky, M., Bottou, L., Gulrajani, I., and Lopez-Paz, D. Invariant risk minimization. ArXiv, abs/1907.02893,
1907 arXiv
-
[2023]
11 A Closer Look at Multimodal Representation Collapse A. Extended Literature Review Missing Modalities: Existing SOTA multimodal fusion approaches do not account for the possibility of missing modalities (Ramachandram & Taylor, 2017; Nagrani et al., 2021; Shi et al., 2021; Ch...
2011
-
[2024]
Lecun, Y ., Bottou, L., Bengio, Y ., and Haffner, P
URL https://arxiv.org/ abs/2312.03096. Lecun, Y ., Bottou, L., Bengio, Y ., and Haffner, P. Gradient- based learning applied to document recognition. Proceed- ings of the IEEE,
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.