Reduce the Transformer rank to improve performance while maintaining LLM without reducing the removal of more than 90% of components in a specific layer

MIT and Microsoft conducted joint research and found that it is possible to improve the task performance of large language models and reduce their size without additional training

In the era of large-scale models, Transformer supports the entire scientific research field with its unique capabilities. Since its introduction, Transformer-based language models (LLM) have demonstrated excellent performance in various tasks. The underlying architecture of Transformer has become the state-of-the-art technology for natural language modeling and reasoning, and has shown strong prospects in fields such as computer vision and reinforcement learning

However, the current Transformer architecture is very large and usually requires a lot of computing resources for training and inference.

Rewrite like this: It makes sense to do this because a Transformer trained with more parameters or data is obviously more capable than other models. However, an increasing number of studies have shown that Transformer-based models and neural networks do not need to retain all adaptation parameters to maintain their learned hypotheses

In general, over-parameterization seems to be very problematic when training models. Helpful, but these models can be heavily pruned before inference. Studies have shown that neural networks can often remove more than 90% of weights without any significant drop in performance. This phenomenon has triggered researchers’ interest in pruning strategies that help model reasoning

Researchers from MIT and Microsoft wrote in the paper "The Truth is in There: Improving Reasoning in Language Models with Layer- Selective Rank Reduction made a surprising finding that careful pruning on specific layers of the Transformer model can significantly improve the performance of the model on certain tasks.

• Please click the following link to view the paper: https://arxiv.org/pdf/2312.13558.pdf

• Paper home page: https://pratyushasharma.github.io/laser/

The study calls this simple intervention LASER (Layer Selective Rank Reduction), The performance of LLM is significantly improved by selectively reducing the high-order components of the learning weight matrix of specific layers in the Transformer model through singular value decomposition. This operation can be performed after the model training is completed without additional parameters or data

During the operation, the reduction of weights is performed in the model-specific weight matrix and layer. The study also found that many similar matrices can be significantly reduced in weight, and typically no performance degradation is observed until more than 90% of the components are removed. The study also found that these reductions can significantly improve accuracy, This finding appears to be not limited to natural language, with performance improvements also found in reinforcement learning.

In addition, this research attempts to infer what is stored in higher-order components so that it can be deleted to improve performance. The study found that LASER answered the correct questions, but before the intervention, the original model mainly responded with high-frequency words (such as "the", "of", etc.), which were not even the same semantic type as the correct answer, and also That is to say, these components will cause the model to generate some irrelevant high-frequency words without intervention.

However, by performing a certain degree of rank reduction, the model's answer can be transformed into correct.

To understand this, the study also explores what the remaining components individually encode, using only their higher-order singular vectors to approximate the weight matrix. It was found that these components described different responses or common high-frequency words in the same semantic category as the correct answer.

These results suggest that when noisy higher-order components are combined with lower-order components, their conflicting responses produce an average answer that may be incorrect. Figure 1 provides a visual representation of the Transformer architecture and the procedure followed by LASER. Here, the weight matrix of a specific layer of multilayer perceptron (MLP) is replaced by its low-rank approximation.

LASER OVERVIEW

Provides a detailed introduction to LASER intervention. A single-step LASER intervention is defined by the triplet (τ, ℓ, ρ), which contains the parameter τ, the number of layers ℓ and the reduced rank ρ. Together these values ​​describe the matrices to be replaced by their low-rank approximations, and the degree of approximation. Researchers classify the types of matrices they will intervene on based on parameter types

Researchers focus on matrices in W = {W_q, W_k, W_v, W_o, U_in, U_out}, which consists of MLP and attention layers composed of matrices in . The number of strata represents the stratum of researcher intervention (the first stratum is indexed starting from 0). For example, Llama-2 has 32 layers, so ℓ ∈ {0, 1, 2,・・・31}.

Ultimately, ρ ∈ [0, 1) describes which part of the maximum rank should be preserved when making low-rank approximations. For example, assuming

, the maximum rank of the matrix is ​​d. The researchers replaced it with the ⌊ρ・d⌋- approximation.

Figure 1 below is an example of LASER. In this figure, τ = U_in and ℓ = L represent updating the weight matrix of the first layer of MLP in the Transformer block of the L^th layer. Another parameter controls k in the rank-k approximation.

LASER can restrict the flow of certain information in a network and unexpectedly yield significant performance benefits. These interventions can also be easily combined, such that a set of interventions can be applied in any order.

The LASER method is simply a search for such interventions, modified to deliver the greatest benefit. However, there are many other ways to combine these interventions, which is a direction for future work.

In order to keep the original meaning unchanged, the content needs to be rewritten into Chinese. There is no need to appear the original sentence

In the experimental part, the researcher used the GPT-J model pre-trained on the PILE data set. The number of layers of the model is 27 and the parameters are 6 billion. The model's behavior is then evaluated on the CounterFact dataset, which contains samples of (topic, relation, and answer) triples, with three paraphrase prompts provided for each question.

The first is the analysis of the GPT-J model on the CounterFact data set. Figure 2 below shows the impact on the classification loss of the dataset as a result of applying different amounts of rank reduction to each matrix in the Transformer architecture. Each of the Transformer layers consists of a two-layer small MLP, with the input and output matrices shown separately. Different colors represent different percentages of removed components.

Regarding improving the accuracy and robustness of interpretation, as shown in Figure 2 above and Table 1 below, the researchers found that when performing rank reduction on a single layer, GPT The factual accuracy of the -J model on the CounterFact dataset increased from 13.1% to 24.0%. It is important to note that these improvements are only the result of rank reduction and do not involve any further training or fine-tuning of the model.

#Which facts will be restored in the data set through rank reduction has become a concern for researchers. The researchers found that the fact of recovery through rank reduction rarely appeared in the data, as shown in Figure 3

What do higher-order components store? Researchers use high-order components to approximate the final weight matrix. Unlike LASER, they do not use low-order components for approximation, as shown in Figure 5(a). They measured the average cosine similarity between the true and predicted answers when approximating the matrix using different numbers of higher-order components, as shown in Figure 5(b)

Finally, the researchers evaluated the generalizability of their findings to three different LLMs on multiple language understanding tasks. For each task, they evaluated the model's performance by generating three metrics: accuracy, classification accuracy, and loss. As shown in Table 1 above, even if the rank reduction is large, it will not cause the model accuracy to decrease, but it can improve the model performance.

The above is the detailed content of Reduce the Transformer rank to improve performance while maintaining LLM without reducing the removal of more than 90% of components in a specific layer. For more information, please follow other related articles on the PHP Chinese website!