BrainBits: How Much of the Brain are Generative Reconstruction Methods Using?

We thank the reviewer for their time and extensive feedback. We answer questions and clarify a few points below.

• The paper suggests that better reconstruction may be achieved with less brain signal input, yet all brain signals are utilized by the reconstruction models.

  • Yes, while all brain signals are available at input, the dimension of the bottleneck puts a hard upper bound on the amount of information available at reconstruction time. Notably, increasing the size of this bottleneck leads to better performance, from which we infer that these bounds are tight.
  • We have preliminary analysis of which parts of the brain are likely important (Figure 5), but truly selecting inputs by dropout is left for future work. This procedure is complicated by differences between subjects and significant computational costs, since dropout is a combinatorially difficult problem.

• Only diffusion-based models are employed for decoding images. The relatively favorable results may stem from well-trained diffusion models capable of operating with minimal semantics. Diffusion models exhibit significant randomness in image generation, enabling the selection of images resembling ground truth stimuli.

  • We do not allow selecting images resembling the ground truth stimuli. We agree with the reviewer that doing so would be highly problematic. Instead, models get one chance to produce an image: one image for each of the 1000 stimuli. Then the average performance is computed. The variance on that average performance is small, since over so many images the stochastic nature of the diffusion models averages out.
  • Regarding the power of diffusion models, we agree entirely! Our entire point is that these models operate with minimal semantics, rather than with extremely rich representations needed to actually encode the entire image. This is not a particular property of diffusion models, it is instead a property of any large model trained on data that is similar to the test data shown to subjects.

• The performance of brain-to-image generation remains unsatisfactory.

  • I think we’re in agreement here! The main thrust of our argument is this: prior work has claimed that brain-to-image reconstruction is satisfactory to some degree, according to existing metrics.
  • And whatever flaws these images have is up for discussion, but our experiments show that you can achieve the same performance on those metrics with much less information from the brain. This is so little information that clearly something must be missing from the images. Yet this point has not been made in the literature and it is critical for further progress.

• The number of samples for different subjects is limited.

  • Our paper should be thought of as a meta-study over the field of work in brain-to-image reconstruction. To this end, we consider those subjects and samples used by existing works.

• Differences in temporal and spatial resolution as well as signal characteristics among recorded signals pose uncertainties about BrainBits' adaptability.

  • Our method is very general: and simply entails adding a restriction on information flow through the reconstruction pipeline. This is applicable, no matter the modality.

• Authors claim reconstruction may require same or less signal from the brain, while also acknowledging the scarcity of open neuroscience datasets.

  • To clarify, there is a distinction between the signal available in a single brain reading and the data available in the training dataset. The first is relevant to the quality of reconstruction, the second is relevant to the ease of training a reconstruction method.

• Regions depicted in Figure 5 are not clearly labeled

  • We created a legend with labeled regions (see additional PDF) that we will add to the main text.

• The effective dimensionality in Figure 4 lacks explicit clarification.

  • Lines 201 through 204 explain the effective dimensionality method used. We will update the caption to include this.

• Text occlusion in both Figure 3 and Figure 11.

  • Thanks! We will add labels, clarification in the caption for Figure 4, and we will fix the spacing in Figures 3 and 11

• What meaning do lines 4-7 convey?

  • The reasoning is this: currently only measuring the quality of the reconstructed images allows for models to improve for reasons that do not involve extracting more signal for the brain. Lines 4-7 simply list a few of these possible reasons. The purpose of our work is to introduce a metric that cannot be “gamed” in such ways.

• Experimenting with other signal types or generative models?

  • The overwhelming number of recent state of the art publications in image reconstruction use fMRI enabled by the NSD dataset. Since this is by far the most common paradigm, we decided to adopt it. There is nothing about our method that is signal-specific. Nor is there anything that is specific to diffusion generative models. Given the state of the field, we chose the most representative methods.

• Do different voxels get activated for different stimulus images within a subject?

  • All of our learned maps and results are per subject. This avoids this problem.

• Can BrainBits operate with fewer brain signals as input? If not, what specific contributions does this paper offer to the community?

  • We are not sure if we understand the question. BrainBits is not a reconstruction method in itself. BrainBits is a method for measuring the dimensionality of the neural recordings required to perform reconstruction. And it is applicable, no matter the input size. It addresses a key problem: large networks can reconstruct images well with little information by exploiting the similarity between their massive training sets and relatively restricted test sets. Without BrainBits, one cannot disentangle why a model is performing better. Is it explaining more of the brain or taking more advantage of its priors? We show that these priors can be extreme.

We appreciate the reviewer's time and good questions, which helped us clarify our interpretations. We're glad that the reviewer found our paper important to the field and easy to follow. Below, we answer the posted questions.

• Can't the range be set according to data dynamic range?
  • The inset is to emphasize the fact that for the text decoding case, reconstructed text is still very close to chance in an absolute sense. We will enlarge the label text for legibility and include a larger plot in the supplement.
• Why do authors claim that the measured expansion is not "meaningful"? What would a meaningful expansion look like?
  • In the 5 vs 50 case we meant to point out that no new significant cluster areas of activity emerge. You are correct in that “meaningfully expand” is not very precise, we will update the language in our paper to better describe our observations. We merely meant to observe that the most salient regions remained the brightest across all scales, and few new islands of activity were highlighted for the larger bottlenecks.
• Figure 11 has cut-out subplot titles that are illegible
  • Thanks for the catch! We’ll fix the spacing. To clarify, the rows of that figure were in subject ID order: 1, 2, 5, 7.
• What are the implications of neural coding redundancy on the analysis presented in Fig. 4 where it is shown that a small fraction of the bottleneck dimension are effectively used by the models? Could it be that indeed the underlying neural code is highly redundant hence the model can make effective use of neural signals by using a much lower dimensionality?
  • For comparison, for brain-diffuser, the average effective dim of the fMRI inputs is 2,257. For Takagi et al, the average is 4,485 (see additional PDF). This suggests that the underlying neural code is far higher dimensional than our bottleneck size, despite any redundancies that may exist.
• Authors propose that each method should report a reconstruction ceiling. How would they suggest to approach this problem in general?
  • The very simplest version of incorporating a ceiling, would be to always show the evaluation metrics, as they are with the ground truth inputs. For example, we show the complete scale from 0.0 to 1.0 on Figure 3c. We argue that looking at the distance to this simple ceiling is an important part of gauging whether reconstruction results can be called successful or not. A slightly more sophisticated ceiling involves using the ground truth image latents as conditioning information for the diffusion model. This is what we do in Figure 3a-b. Both of these procedures should always be feasible.

We thank the reviewer for recognizing the importance of our work and for their time and feedback.

• The main weakness is that the bottleneck introduced in the paper does not definitely answer how much of the brain signal has been exploited in the process. Indeed, the MLP projecting the fMRI data to the bottleneck actually learns to identify the most important features within the brain data, in order to obtain the best reconstruction. Much like a VAE, it learns to compress information as efficiently as possible in order to have the best performance possible. Thus, the compression ratio, even if it is generally very high in the authors' results, is also a result of the best effort of the MLP to identify the best features. This crucial point kind of defeats the point of the paper: isolating how much of the brain signal was extracted and used. However, given the size of the compression ratio (300 in the case of BrainDiffusers for instance), I still believe that the authors rightfully identified that not all of the brain signal is used, and that their research is on the right track. However this point could be further improved and disentangled.

  • First, to clarify, we use a single linear layer, not an MLP, for projection. The nice thing about using a single linear mapping is that we can be fairly confident that no extra expressive power has been added to the reconstruction method. A single linear layer has much less representational power than a VAE. Of course, it is likely that better compressions may be possible with more powerful projection mappings, but these would be less interpretable as the reviewer correctly points out.

We greatly appreciate your time in reviewing our paper. Please let us know if our response helped clear anything up, or if there is anything else you would like answered!

Thank you for your time and feedback. We're glad you found our work timely, sound, and convincing. We appreciate your questions, which have helped us sharpen our descriptions.

• What is the effective dimensionality of the actual neural activations? How can we know that the plateau in performance at a small bottleneck size reflects the model ignoring usable information in the neural measurements, versus exhausting the usable information available in the neural measurements?

  • The reviewer asks a good question! What is the dimensionality of the input fMRI signal and are we simply recovering this number? For brain-diffuser, the average effective dim of the inputs is 2,257. For Takagi et al, the average is 4,485. In comparison, 50 is small. (See additional PDF). We will add this figure to the appendix and note this computation in the main text.
  • We are glad the reviewer found our paper well motivated. We would like to further add that our main contention is not simply that the amount of needed brain data is small in an absolute sense, but that, as a field, we should always use a metric that is sensitive to the amount of brain data used for reconstruction, so as not to be misled by methods that improve on the image prior without improving on the extraction of neural signal.

• Under the hypothesis that the prior in generative models is strongly contributing to reconstruction performance, it is odd to me that higher-level visual areas don't seem to be used by the reconstruction pipeline, even at higher bottleneck sizes: given that their activation is highly informative regarding object category, and given that these models presumably have a strong prior for how objects tend to look, I wouldn't have predicted this. It would be interesting (though perhaps too much for the scope of this paper) to restrict the pipeline to particular sectors of the visual system, and examine how this affects performance.

  • We suspect that higher level areas aren’t being used in part because they are redundant with early areas. Representations in early areas are retinotopic and potentially simpler than those in later areas which may be neither.
  • And yes! This is exactly what we aim to do in subsequent work and the main other application of our method: probing the information available in the decodings conditioned on different areas of the brain. We plan to automate this by having models look at the resulting images and then quantify the information that is being decoded: like albedo, lighting, shape, texture, class, relationships, etc.