Scaling laws for language encoding models in fMRI

Summary

The authors evaluate the effect of LM scale (in terms of N parameters) and fMRI dataset size on the performance of downstream encoding models, trained to predict the activity of individual voxels across the brain as a function of the input those participants received during a scanning session. They explore both the OPT and LLaMa models as pure text based LMs and the HuBERT, WavLM, and Whisper models as acoustic models, across a number of sizes. The authors find that fMRI encoding performance can be improved linearly via logarithmic increases in model scale, up to ~30B for text-based models, and the asymptote has not yet been reached for acoustic models. Increases in the amount of fMRI data included per subject also yield dramatic improvements. They also attempt to characterize what additional encoding performance is yielded by the use of acoustic models over text-based ones and find improvements only in auditory areas but not in higher-level associative areas. Such results propose for increased focus on “deep” neuroscience data collection and scaling up the computational resources utilized for encoding analyses.

Strengths

The work explores a novel and relevant axis for brain encoding analyses and is appropriately contextualized via discussion of related work. The submission is technically sound and most claims are well supported. The submission is clearly written and well organized. The results are relevant and important and will hopefully encourage a “scaling up” of neural encoding analyses.

Weaknesses

The sole evaluation of LM scale as a function of N parameters is reductive and should be addressed. There is a lack or under-specification of uncertainty quantification. See below for more detailed comments.

Questions

General: Aside from LM model size (N parameters) a crucial component in predicting LM performance is the size of the pretraining dataset (e.g., OPT’s 180B tokens vs. LLaMa’s 1.4T tokens). Given that the authors observe differences between the architecturally similar LLaMa and OPT model families, but these models differ substantially in pretraining dataset size, this is worth mentioning. Relatedly, for the combined analysis of the LLaMa and OPT models, a more unified axis along which to evaluate the encoding performance of these models might be their mean perplexity across the evaluated podcast stories dataset. I strongly encourage the authors to consider this visualization as a parallel to 1a. It is both likely to be informative, and also proposes a more functional view of the problem space. One would be unlikely to encourage the use of a larger but undertrained 30B model over a smaller 6B model trained on more data with lower perplexity for an fMRI encoding task, particularly due to the author-mentioned advantages of more compressed representations and suggestion that “a careful balance must be struck between model size and model efficacy in order to maximize encoding performance”. Error bars: What do the error bars in figure 1b and 1e represent? (stdev, se, x% ci, etc.?) No uncertainty quantification is included for any other figures. Given that “Encoding model performance for a given layer was computed as the average voxelwise performance of that layer’s hidden states across of all of cortex for all of our 3 subjects”, there would be a number of ways to include uncertainty quantification at this top level even without rerunning any models. Minor: increase resolution of figures 1a,1b,1d,1e,3b,3c.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

4 excellent

Contribution

4 excellent

Limitations

The limitations of the work are appropriately addressed.

Summary

Researchers compared the effectiveness of larger open-source language models, such as those from the OPT and LLaMA families, in predicting brain responses recorded using fMRI. They found that brain prediction performance improves logarithmically with model size, with a 15% increase in encoding performance as model size increases. Similar improvements were observed when scaling the size of the fMRI training set. The study also explored the scaling of acoustic encoding models and found comparable improvements. The analysis suggests that increasing the scale of both models and data will lead to highly effective models of language processing in the brain, enabling better scientific understanding and applications such as decoding.

Strengths

- Presents a novel empirical observation of scaling laws for language and audio encoding models in fMRI. - They show log linear scaling in brain prediction, encoding performances in both language and acoustic domains. - They also show when scaling the size of the fMRI data, the performances increase log-linearly.

Weaknesses

Most of my concerns are addressed in the questions section. One concern I have is that the code is not publicly available. If the authors published the code it would be greatly appreciated.

Questions

- Wernicke's area is often paired and observed alongside Broca's area. Did the authors consider observing the Wernicke's area for both auditory and language encoding models? - In section 3.4 figure 3a, the "room for improvement" seems to be slightly more prominent in the right hemisphere. Please note that this observation is purely from the figure. Have the authors explored why this may be the case? - To observe the scaling laws in increasing data size, the authors withheld different amounts of the training data. Did the authors consider augmenting the data in other ways?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The authors adequately addressed the limitations of the paper.

Summary

The authors study how the scaling of large language models produce accurate features to predict human brain fMRI activity. They report scaling-like laws for fMRI encoding models – models with more parameters tend to more accurately predict fMRI activity (as measured with the Pearson correlation of the encoding model predictions and fMRI time series).

Strengths

I found the paper to be of general interest, and the methods to be sound. The paper suggests that there is a utility to using modern large ML models as producing features for fMRI encoding models.

Weaknesses

It is hard to determine whether this result should be expected, and what to make of its scientific significance (beyond an engineering feat). On one hand, one would expect that models that are larger (and therefore better at modeling language/sound data) should produce more useful features for language/sound data. On the other hand, these large models process language/sound data in ways that are almost certainly distinct from how the brain processes information (e.g., the brain doesn’t appear to merely do next-word prediction). So what does one make of this discrepancy? Do the authors believe that the brain mimics the computational processing of these models? And what is the scientific significance of using a model – that presumably looks nothing like the brain – to build better models of the brain? How valuable is using percent correlation change as a metric for how well the encoding models perform relative to, for example, R^2? Is it truly useful to benchmark or establish ‘scaling laws’ based on the average encoding performance across all fMRI voxels, rather than identify one or two key regions (or the variability across regions)? What might be interesting to evaluate is whether these scaling laws only apply to a subset of voxels/brain regions, since, presumably, how well a voxel/region follows a scaling law is likely dependent on how well a brain region’s function matches with the model type.

Questions

For an ‘fMRI encoding model practitioner’, it would be helpful if there were concrete ‘recommendations’ for what models are most appropriate for language/audio modeling. For example – and this is a minor point – does it matter which layers one uses to produce the regression model features? If possible, it would be interesting to see a surface map that shows which voxels/regions follow a scaling law for the audio v language models, respectively. How dependent is the scaling law on computing the average (across voxels) for each of the maps per model? Can the authors also report R^2 for their encoding models benchmark? What does it mean (scientifically) that the model that best produces regression features for brain encoding models look nothing like the brain? Are there conclusions one should draw from this? Or is this just a coincidence? Might there be any correspondence between layers in an LLM and different brain regions?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

The limitations are mostly in terms of the scientific utility of claiming that there are scaling laws for fMRI. I doubt the authors intend this interpretation, but the title seems to hint at the possibility that if one achieves enough scale with language models, one might be able to perfectly predict brain activity (and therefore model the brain). This seems implausible (and likely not intended), but some discussion around this would be helpful.

Summary

Previous works have demonstrated that activations from Language Models fit, to some extent, brain activations in participants listening to audio texts. Here, the authors examine how the fit is influenced by model size (number of parameters) and the training dataset size. The main observation is that performance scales log-linearly with model size (until 30B parameters) and with fMRI training set size (from 1 to 100 stories).

Strengths

The main originality of the work is to explore language models that are much bigger than the ones that had been used in previous works of this type (as well as recent, large, acoustic models). Not only does this represent a technical feat, but it also advances knowledge on the topic, as it was not clear if the representations constructed by larger models would necessarily converge towards more and more brain like representations when the models become bigger. Data presented here show that there was still a lot of room for improvment from GPT-2 like models. The raw performance of the best model (OPT-30B), presented in Fig.2 is quite impressive and, along with the analysis of the effect offmri training set size, buttresses the authors' claim of the usefulness of using large within subject dataset. Fig.3b which compares performance to estimated ceilings is very interesting theoretically as it suggest that the models model some regions better than others.

Weaknesses

As the authors mention (at line 300), when models grow in size, the regression problem may become ill conditioned. Would there be any strategy to remedy this problem?

Questions

Figure 1 shows the increase in performance averaged _across the whole cortex_. Have you checked whether this profile is similar in different regions (e.g. auditory cortex, core language region, and higher-level regions like precuneus and medial prefrontal)? Were the (OPT) models of different sizes trained with the same language corpora (fixed size)?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

The limitations were adequately discussed.

Summary

The authors of this paper delve into the investigation of the scaling law between the performance of predicting brain activity (measured using BOLD) and the number of parameters in large language models, coupled with the amount of training data for the linear readout for retrained LLM used in AI. This pursuit of exploring a log-linear scaling law relationship is an interesting and valuable. It holds promise in providing crucial information for paving a path forward in developing better functional models of brain activity, which could be instrumental in understanding brain computations. The technical dimensions of the paper are well-executed. The authors have displayed a solid grasp of the subject matter and have employed appropriate methodologies to carry out their research. However, I have some reservations about the validity of their claims concerning the non-saturation of the scaling law for the number of parameters in large models. Specifically, Figure 1, which is central to their argument, seems to present evidence contrary to their claim. They acknowledge a visible saturation but attribute this observation to memory limitations in the way they fit the linear readout (i.e. they select one layer). However, unless they successfully demonstrate that the saturation does not occur when a more sophisticated linear readout is utilized (for example, from all layers for larger models), it is not justified for them to assert that the predictive models demonstrate scaling as a function of parameter size. Thus, the authors need to refine their claims or provide additional analysis to convincingly demonstrate that the saturation effect they observed is truly an artifact of their methodological limitations, not an inherent feature of the models they are studying. Without such demonstrations, the claim about the non-saturation of the scaling law remains unsubstantiated. In conclusion, while the paper tackles an intriguing concept and showcases strong technical elements, the authors need to strengthen their assertions through concrete evidence or refinement of their conclusions. This additional effort will greatly enhance the paper's overall quality and the validity of their findings, adding substantial value to the body of knowledge in the field.

Strengths

The question of how far scaling deep learning models can continue to improve neural prediction is an important question in the quest to build more accurate models of the brain.

Weaknesses

More work is needed to demosncate that the current LM are not saturating the performance of the neural prediction task. This is very critical since the main contribution of this work is that even the largest LM do not saturate the performance.

Questions

The authors need to demonstrate using other linear readouts that there is not saturation of performance in the neural prediction task.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

yes

Comments addressed

Thanks for addressing this review and for supplying new content. > You are correct, we should have included this detail. We will mention this in an appropriate place in the camera-ready version of this paper. > We apologize for the oversight. These error bars represent standard error. This will be clarified in the camera-ready version. > Apologies. We intend to revise this for the final camera-ready version. Great. > To address your concern, we have computed the mean perplexity across one of our test stories for all but our largest model (which we cannot rerun due to limitations in our computational budget). We would like to note that the perplexities are not directly comparable across the different model families as they employ different tokenization schemes. The table is included below and will be included in some form in the final version. Ah, yes, you are right about tokenization, but thanks for producing this table. Even with that caveat, this is still an informative supplement. > We chose to omit error bars in Figure 1a and 1d as the only sources of uncertainty quantification would be along the voxel axis or the subject axis. Quantifying uncertainty along the voxel axis would be more misleading than helpful as some voxels are inherently less responsive to language. Quantifying uncertainty along the subject axis is also not necessary here as we plot the individual performance for each subject as well as the mean. Figure 2 similarly shows uncertainty estimation by plotting all individual time courses of the 10 test story repeats. We will add to the final camera-ready version SNR-normalized subject-axis uncertainty estimation to Figures 1c and 1f, which should give a sense of how the shape of the layerwise encoding performance curve varies from subject to subject. This uncertainty estimation can be seen in Figure 2 of the rebuttal PDF. Okay, all these decisions are reasonable and the proposed actions look good. Thanks!