How does Architecture Influence the Base Capabilities of Pre-trained Language Models? A Case Study Based on FFN-Wider and MoE Transformers

Pre-trained language models have been proven to possess strong base capabilities, which not only excel in in-distribution language modeling but also show powerful abilities in out-of-distribution language modeling, transfer learning and few-shot learning. Unlike existing work focusing on the influence of scale on base capabilities, our work examines the influence of architecture on those. Specifically, our concern is: How does architecture influence the base capabilities of pre-trained language models? In this work, we attempt to explain and reverse the decline in base capabilities caused by the architecture of FFN-Wider Transformers, seeking to provide some insights. Through analysis, we found the contribution ratio of Multi-Head Attention (a combination function) to pre-trained language modeling is a key factor affecting base capabilities. FFN-Wider Transformers reduce the contribution ratio of this combination function, leading to a decline in base capabilities. We confirmed this by experiments and proposed Combination Enhanced Architecture (CEA) to address the decline in base capabilities of such models. Significantly, we extended our explanation and CEA to Mixture of Experts (MoE) Transformers. We successfully achieved significant improvements in base capabilities on a 14B parameter MoE model, demonstrating the practical application value of our work. This also indicates that our analysis has a certain guiding significance for architecture analysis, architecture improvement and architecture design.

Paper

References (35)

Scroll for more · 23 remaining

Similar papers

Peer review

Reviewer nte26/10 · confidence 3/52024-07-08

Summary

This study investigates the influence of architecture on pre-trained language models' base capabilities. It reveals that the contribution ratio of Multi-Head Attention to pre-trained language modeling affects base capabilities. FFN-Wider Transformers reduce this ratio, leading to a decline in base capabilities. The researchers propose a Combination Enhanced Architecture (CEA) to address this issue. They also extend this to Mixture of Experts (MoE) transformers, achieving significant improvements in base capabilities.

Strengths

+ The paper shifts the focus from the commonly studied impact of scale on pre-trained language models to the impact of architecture. + The exploration of the influence of using wider FFN layer is interesting, and should impact further researches. + The proposal of CEA as a solution to the decline in base capabilities shows a proactive approach to solving the identified problem. + The findings are confirmed through extensive experiments, adding credibility to the claims.

Weaknesses

- The description of CEA is not detailed enough.

Questions

In general, I think the discovery of the paper is interesting, and I don't have many questions. Just some minor issues: - What is the definition of mutual information? Give some introduction. - How is the pre-training performance of the models? - How is the structure of CEA? It is not clear to me after reading Section 6.

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors address it in Section 8.

Reviewer QyvJ6/10 · confidence 3/52024-07-09

Summary

The paper studies the contribution raion of FFN and MHA layer in transformers and its effect on out-of-distribution performance. It finds that the wider FFN layer will increase its contribution ratio and lower the OOD performance. Lastly, the paper proposes a new architecture that moves part of the FFN layer to MHA and shows the CEA can improve the baseline models' OOD ability.

Strengths

- The paper analyzes the contribution ratio with new methods that evaluate Mutual Information and token prediction accuracy of each layer's output. - The proposed CEA method has improved the MoE transformer in various tasks and datasets with the same parameter scale, showing its effectiveness in increasing OOD ability.

Weaknesses

- The paper does not give a convincing explanation of why the study aligns pre-trained performance with different parameter scales. A larger scale model may not fully converge when it has a similar training loss to a smaller model. It is also not known whether the proposed CEA will harm in-domain performance. - The MHA module in the vanilla transformer also has linear layers for Q, K, V, and after self-attention. I think they can be seen as inner-FFNs that will transform the combination function and make it not necessary to introduce an extra inner-FFN layer. - Tables 1 and 2 show that BERT w/ CEA performs worse than vanilla in GLUE and SGLUE and GPT w/ CEA perform worse in all datasets.

Questions

See weaknesses 1 and 2. Others: - Figure 6 lacks a legend to explain the figure elements, including the blue line, orange bar, blue and orange dotted lines.

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors addressed the limitations.

Reviewer Rvop4/10 · confidence 3/52024-07-13

Summary

This paper examines how the architecture of a transformer model influences its base capabilities, such as out-of-distribution tasks, transfer learning, and few-shot learning. Specifically, it explores the effects of replacing the feed-forward network (FFN) with a wider FFN (FFN-wide) in various parts of the architecture. Initially, the authors find that replacing the FFN (referred to as the transformation function) after the attention layer (referred to as the combination function) with FFN-wide leads to worse performance. They analyze this impact by measuring the contribution of the combination function using techniques such as mutual information and token prediction. The results indicate that performance deteriorates in most cases when the contribution of the combination function decreases. The authors then devise multiple architectures with different width ratios of FFN-wide in the combination and transformation functions. They select the best ratio model architecture, termed CEA (Combination-Enhanced Architecture). This CEA architecture is subsequently used in a Mixture of Experts (MoE) model, demonstrating improvements over the non-CEA MoE model.

Strengths

- The paper addresses an interesting problem by measuring the impact of transformer architecture on the downstream performance of the model, which is crucial for optimizing and understanding transformer models. - Attributing performance to the inner workings of a neural network, particularly the contribution of specific components, and analyzing this contribution using techniques like mutual information is both novel and insightful. - It is also interesting to see the impact on two different kinds of model like Bert and GPT and see how the optimal width for the two types of models are different. - The application of CEA to a mixture of experts models definitely demonstrates the practical benefits of this paper.

Weaknesses

- The performance improvements for the different variations of the architectures seem minor, and the absence of standard deviations makes it difficult to assess the robustness of the results. - For both BERT and GPT, even after a thorough architecture search, the performance is at best similar to the vanilla models, which is not very promising. - Similarly for the MoE experiment, it will useful to add the standard deviation of the results. Additionally, since the training loss is lower for MoE with CEA compared to vanilla MoE, it is hard to determine whether the improvements are due to the architecture's inductive bias or just the lower training loss. Similar to the rest of the paper, please also report numbers for this experiments by keeps the same training performance level. Overall, the paper is interesting, but the performance values being so close and the lack of variance measurement do not give me much confidence in the results. Minor: In Table 1 and Table 2, the results for the proposed approach are bolded rather than the best approach for that experiment. In most figures, the y-axis does not start at 0, which can give readers an inaccurate representation of the data.

Questions

- What does it mean for the width ratio to be 0% and 100%? Does this imply that the FFN is removed in these cases? - What does it mean that FFN-wider models with CEA are not able to beat vanilla models in some cases in Table 1 and Table 2? Isn't the vanilla model one of the ratios explored? If not, would it make more sense to explore the architectures on some other dimensions? - Please indicate whether lower or higher values are better for all the tables and figures in the paper.

Rating

4

Confidence

3

Soundness

2

Presentation

2

Contribution

3

Limitations

The authors can also address the limitations of exploring only a narrow subset of architectures and discuss whether any parts of the proposed methodology can be extended to other aspects of transformer architectures. Additionally, all experiments were performed on smaller-scale models, so the results may not generalize to larger models.

Authorsrebuttal2024-08-14

Thank you very much for reviewing our work! We have addressed some of the concerns you raised in our rebuttal. Specifically, for the issues you were particularly concerned about, such as the fairness of the MoE experiment comparisons and the standard deviation of the experimental results, we have provided more detailed results (in the global response PDF above). We strongly believe that these additional results will effectively address your concerns, especially regarding the fairness of the MoE experiment comparisons, which seems to have led to some misunderstanding. If time permits, we kindly ask you to review our rebuttal. We look forward to your feedback. Thank you!

Reviewer xAuN7/10 · confidence 4/52024-07-13

Summary

The paper examines the influence of architecture on the base capabilities (OOD, transfer learning, few-short learning) of large language models. The main focus is on FFN-Wider transformers and understanding why they have poorer base capabilities compared to vanilla transformers. The contibution ratio of multihead attention to pretrained language modeling is found to be a key factor affective base capabilities. Based on this observation, a Combination Enhanced Architecture is proposed and also extended to mixture of experts transformers. The analytical study is backed by experimental evaluation successfully achieving significant improvements in base capabilities of the 14B parameter MoE model.

Strengths

+ While most existing study of large models has focussed on the impact of scale, this work is a significant effort in understanding the influence of architecture. + The empirical findings are supported by a more deeper interpretation of the underlying mechanisms of these influences. + The proposed CAA wherein the wider FFN is split into adjustable two parts - outer-FFN which stays as a transformation function, and an inner-FFN which is relocated within the MHA layer for enhancing the combination function is interesting in its own right. + The extension to MoE transformers and the experimental demonstration with 14B parameter GPT architecture MoE model is very interesting.

Weaknesses

- Some of the findings appear to be obvious (perhaps, in hindsight) - as the contribution ratio of the MHA layer increases, the base capabilities also improve. It would have been interesting to also find some non-intuitive influences. - Overall the paper's analysis is much narrow in scope than the title of the paper. Nonetheless, it is interesting and useful study. - The language and presentation of the paper can be improved. It will be useful to have the grammar issues fixed in the paper before the final version.

Questions

- Some choices of parameters such as chosing intermediate dimension to 32d instead of say 8 or 16d can be better explained. - Was the width adjustment limited to Outer-FFN?

Rating

7

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

The reviewer does not except any negative societal impact of this work.

Reviewer QyvJ2024-08-13

Thanks for the response, it addresses most of my concerns. I still have some concerns about the results of BERT/GPT w/ CEA. I understand that the FFN-Wider performs worse than the vanilla model, but I think w/ CEA should at least achieve similar results to the vanilla model. Otherwise, there is no reason to use CEA rather than vanilla. Maybe authors can show that the CEA can outperform vanilla with the same parameter scale and not align pre-training performance. However, the results of MoE are convincing. I think this is a good paper and will raise my rating.

Reviewer nte22024-08-14

Thank you for the rebuttal! It addresses my concerns and I will keep my rating.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC