acceptodds
Under review as a conference paper at ICLR 2027

What Makes the Weights of Implicit Neural Representations Classifiable?

Abstract

Implicit neural representations (INRs) store an image in the parameters of a small network, and recent work classifies images by reading these parameters. The Meta Weight Transformer (MWT) made such classifiers far more accurate by learning, jointly with a Transformer classifier, how sinusoidal INRs (SIRENs) are fitted. Earlier work relates accuracy to measures of class structure in the parameters, but such measures show what can be read from the parameters, not what a classifier uses. In this paper, we study whether class structure reflects how well MWT classifies, and which inputs its classifier relies on. We first compare nearest-neighbor and nearest-centroid readouts of this structure with the trained classifiers across six MWT training configurations. The readouts are far less accurate and do not rank the configurations by accuracy. Next, we change the inputs of the trained classifiers without retraining them, and find that the classifiers depend heavily on the per-neuron biases, which make up less than 0.4% of their input values. Replacing the biases by their layer means lowers accuracy by 16.3 points on average, whereas the same change to a weight column of the same size lowers it by only 0.3. This dependence centers on a low-dimensional variation of the biases across images, one that closely matches coarse color and brightness statistics. Our results show that, in MWT, the fitted parameters are classifiable in part because their biases carry coarse information about the image.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
implicit neural representations, weight space learning, metanetworks, SIREN, interpretability, probing, interventions

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_gSh3OpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

Authors investigate an important and underexplored problem in weight-space learning, which is about understanding how meta-weight transformers (MWTs) perform classification. In the general setting, meta-learning is applied to develop implicit neural representations (INRs) that both represent target images and have parameters that can be processed to classify the represented image. However, it is generally unclear how the weights are organized to jointly enable image classification and representation.

The mechanistic analysis of MWT is the main contribution of the paper. Here, authors conduct several experiments, including modifying and analysing INR biases, which they found to be important for the classification. Authors also modify the MWT method in several ways, slightly improving performance over the baseline.

Critical strengths
  • The investigated problem is important for the field. While several approaches to INR classification exist, it is not well understood how these methods work. In MWT, it is unclear how meta-learning organizes classification related information within the INR weights.

  • Authors perform broad experiments, including PCA based analysis and various interventions on INR weights. Obtained results illustrate that MWT organizes the weigths in a way that support classification (but this is not suprising in some sense).

  • Authors improve the original MWT method. On Imagenette (Table 3), the Composite method achieves accuracy of 62.33 compared to 60.65 for the baseline.

Critical weaknesses
  • The interpretability analysis could be more systematic. The applied interventions are mostly handcrafted (e.g., zeroing biases). The paper would benefit from using methods that estimate parameter importance automatically. The authors mention in the Related Work that gradient based attribution methods have already been used to determine the importance of INR parameters for classification [Janik et al., ECCV 2026]. Given that ECCV was in September, close to the ICLR submission deadline, I would not consider the lack of comparison with this paper an issue for this work. However, using some kind of automatic attribution method for INR parameters would help understand whether biases are important for classification without directly targeting them (do they emerge as important?). Another possibility would be to adapt randomized masking methods, such as RISE (Petsiuk et al., BMVC 2018), from image pixels to groups of INR parameters. In my understanding, many methods developed for black-box classifiers or transformers can be applied here.

  • The methodological novelty is limited. The authors apply established methods (probes, regression, PCA, etc.) to the investigated problem. There is no tool that is specific to INR classification. Also, the main results (bias dependence) are less suprising given the FUNCTA method/paper.

  • Authors do not investigate the WT method (no meta-learning).

  • Aside from the interpretability study, authors introduce an MWT modification named Composite. However, the Composite method includes various modifications (extra embeddings and schedulers for some training parameters), which do not feel connected to the interpretability results presented by authors. Also, the modifications feel more like engineering than a fundamental improvement of the baseline. For instance, the feedback coefficient is changed from 0.001 to 0.1 during training. In the MWT paper, the reported ablation results stated that setting this value to 0.001 (vs 0.01) actually decreases performance (Table 4). As of now, it is difficult to understand which modifications are actually responsible for the improvements without extra ablations and explanations.

  • The pixel sampling fractions reported in Table A2 raise questions about whether the baseline reproduces the original CIFAR-10 training procedure. In the original MWT paper, the sampling fraction for CIFAR-10 was set to 100% (no undersampling requirement for low-resolution images). It is a little bit unclear why the authors decided to apply sampling in this case.

  • To better illustrate the usefulness of various parameters, the authors could train an MWT that only tokenizes the biases in INRs. For original MWT, they could limit the gradient flow from the classification branch to the biases only. Also, it would be interesting to use a SIREN without the biases just to check if the MWT can place the classification related infromation in the weights. The interpretation of bias dependence is incomplete.

  • Authors perform various interventions, such as zeroing the biases. Could authors also include some images showing how the interventions change the images reconstructed by the INRs? This would be a nice extra qualitative result. Alternatively, does the modification of the bias changes the reconstruction PSNR? In general, do these modifications actually impact the reconstructed images, or do the biases perhaps serve as a placeholder for information required for classification, which happen to be related to image statistics?

Questions for authors

My initial recommendation is reject.

I read this study as addressing an important problem, but I consider the analysis to be somehow preliminary (in its current form). The main issue is the lack of attribution methods that would automatically highlight biases as important for the classification. Also, the Composite method includes several modifications, which are not clearly motivated/explained in the text.

Given the rebuttal window, to improve the odds of changing the rating, I would suggest:

  • Use a gradient based method (or RISE like technique) to determine parameter importance scoresfor the baseline. This should be straitforward given a checkpoint on Imagenette/CIFAR-10. Report the importance distributions for weights vs biases.
  • Train a baseline MWT without the biases on Imagenette. Would MWT just place the classification required information in the INR weights? Optionally, you can increase the width of the INR to show that extra parameters do not compensate the lack of the biases.
  • Better explain the Composite method.
  • Clarify the role of biases in image reconstruction.
Reviewer_BVnsOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the Meta Weight Transformer (MWT), analyzing why the implicit neural representation parameters obtained through joint training can be used for classification. The authors compare the class structure read out from the parameter space with actual classification performance, and through input interventions, they find that hidden layer biases—accounting for less than 0.4% of the input—significantly affect classification accuracy, with their variations across images highly concentrated in a low-dimensional subspace related to coarse-grained color and brightness statistics. Furthermore, the paper compares the effects of different classifier architectures and auxiliary training objectives.

Critical strengths
  1. This is an interesting finding that helps to understand how the INR parameters in task training encode useful information.
  2. The core deviation-dependent results is supported by extensive experiments.
Critical weaknesses
  1. After removing the components of predictable colors and the principal components that match the variance, the degradation effect is almost the same, which leads to the unclear specific role of color.
  2. Color-to-bias regression is fitted and evaluated on the same samples, while probe comparisons use different data splits. Architectural variants also change multiple factors simultaneously, making individual contributions difficult to isolate.
Questions for authors

Please refer to weaknesses.

3 comments
AuthorsOpenReview ↗

Thank you for your review

As the discussion period comes to a close, we would like to thank you again for your thoughtful review, and we hope you might have a chance to look over our response and revision plan. Your point comparing color removal with the leading principal components, along with your suggestions on held-out regression, unified probe splits, and the Composite ablation, helped us clarify the specific role of color in the paper. If any question remains, we would be glad to address it, and we would greatly appreciate your feedback.

Reviewer_BVnsOpenReview ↗

Rating after rebuttal

After reading the authors responses to my comments and to the other reviewers, I raise my rating to weak acceptance.

AuthorsOpenReview ↗

Thank you for your responses

Thank you for taking the time to read our responses and raise your rating.

Reviewer_SNhuOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper aims to understand MWT (Meta Weight Training) uses to classify an image when the image is represented by a small neural network. The authors first show that simply grouping similar network parameters do not explain how well the classifier performs. Then, they show that biases matter a lot: although they make up less than 0.4% of the input, if they are replace by the mean, accuracy drops by about 16 percentage points. Changing the same number of weights has almost no effect. They also find that most differences in biases across images can be explained by simple color and brightness statistics, and removing that information causes a large drop in accuracy. The main finding is that the classifier relies partly on basic image information stored in the biases. Improving how the classifier reads the parameters helps, while forcing parameters from the same class closer together hurts performance.

Critical strengths

The paper's claim is easy to understand and follow throughout the paper: The biases encode

Critical weaknesses

A full ablation study is missing, since some of the configurations change attention, pooling, training feedback, and pixel sampling together. Without separate ablations, it hard to track the impact of individual changes.

I think there is a limitation in the image dataset sizes, so far the study only applies to small 10-class datasets, any effort to provide at least a slightly bigger dataset would be appreciated, it would show how good can it scale.

The finding about color statistics and their relationship with biases is interesting, but a control may strengthen the interpretation. For instance, could you recolor the same images while preserving their content, refit the SIRENs, and compare both the resulting biases and classification predictions?

Questions for authors

What do you see as the main implications of these findings for future work? In particular, should bias dependence be exploited to build more efficient classifiers, or reduced to improve robustness? What next experiments can help distinguish these directions?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.