اشترك بالحزمة الذهبية واحصل على وصول غير محدود شمرا أكاديميا
تسجيل مستخدم جديدPruning Redundant Mappings in Transformer Models via Spectral-Normalized Identity Prior
341
0
0.0
(
0
)
اسأل ChatGPT حول البحث
ﻻ يوجد ملخص باللغة العربية
Traditional (unstructured) pruning methods for a Transformer model focus on regularizing the individual weights by penalizing them toward zero. In this work, we explore spectral-normalized identity priors (SNIP), a structured pruning approach that penalizes an entire residual module in a Transformer model toward an identity mapping. Our method identifies and discards unimportant non-linear mappings in the residual connections by applying a thresholding operator on the function norm. It is applicable to any structured module, including a single attention head, an entire attention block, or a feed-forward subnetwork. Furthermore, we introduce spectral normalization to stabilize the distribution of the post-activation values of the Transformer layers, further improving the pruning effectiveness of the proposed methodology. We conduct experiments with BERT on 5 GLUE benchmark tasks to demonstrate that SNIP achieves effective pruning results while maintaining comparable performance. Specifically, we improve the performance over the state-of-the-art by 0.5 to 1.0% on average at 50% compression ratio.
قيم البحث
اقرأ أيضاً
Pruning is an effective method to reduce the memory footprint and computational cost associated with large natural language processing models. However, current approaches either only explore head pruning, which has a limited pruning ratio, or only fo
cus on unstructured pruning, which has negligible effects on the real inference time and/or power consumption. To address these challenges, we develop a novel MultiLevel structured Pruning (MLPruning) framework, which uses three different levels of structured pruning: head pruning, row pruning, and block-wise sparse pruning. We propose using a learnable Top-k threshold, which employs an adaptive regularization to adjust the regularization magnitude adaptively, to select appropriate pruning ratios for different weight matrices. We also propose a two-step pipeline to combine block-wise pruning with head/row pruning to achieve high structured pruning ratios with minimum accuracy degradation. Our empirical results show that for bertbase, with textapprox20% of remaining weights, OURS can achieve an accuracy that is comparable to the full model on QQP/MNLI/squad, with up to textapprox3.69x speedup. Our framework has been open sourced~cite{codebase}.
Multi-head attention, a collection of several attention mechanisms that independently attend to different parts of the input, is the key ingredient in the Transformer. Recent work has shown, however, that a large proportion of the heads in a Transfor
mers multi-head attention mechanism can be safely pruned away without significantly harming the performance of the model; such pruning leads to models that are noticeably smaller and faster in practice. Our work introduces a new head pruning technique that we term differentiable subset pruning. Intuitively, our method learns per-head importance variables and then enforces a user-specified hard constraint on the number of unpruned heads. The importance variables are learned via stochastic gradient descent. We conduct experiments on natural language inference and machine translation; we show that differentiable subset pruning performs comparably or better than previous works while offering precise control of the sparsity level.
Large Transformer models have achieved impressive performance in many natural language tasks. In particular, Transformer based language models have been shown to have great capabilities in encoding factual knowledge in their vast amount of parameters
. While the tasks of improving the memorization and generalization of Transformers have been widely studied, it is not well known how to make transformers forget specific old facts and memorize new ones. In this paper, we propose a new task of emph{explicitly modifying specific factual knowledge in Transformer models while ensuring the model performance does not degrade on the unmodified facts}. This task is useful in many scenarios, such as updating stale knowledge, protecting privacy, and eliminating unintended biases stored in the models. We benchmarked several approaches that provide natural baseline performances on this task. This leads to the discovery of key components of a Transformer model that are especially effective for knowledge modifications. The work also provides insights into the role that different training phases (such as pretraining and fine-tuning) play towards memorization and knowledge modification.
The power prior is a popular tool for constructing informative prior distributions based on historical data. The method consists of raising the likelihood to a discounting factor in order to control the amount of information borrowed from the histori
cal data. It is customary to perform a sensitivity analysis reporting results for a range of values of the discounting factor. However, one often wishes to assign it a prior distribution and estimate it jointly with the parameters, which in turn necessitates the computation of a normalising constant. In this paper we are concerned with how to recycle computations from a sensitivity analysis in order to approximately sample from joint posterior of the parameters and the discounting factor. We first show a few important properties of the normalising constant and then use these results to motivate a bisection-type algorithm for computing it on a fixed budget of evaluations. We give a large array of illustrations and discuss cases where the normalising constant is known in closed-form and where it is not. We show that the proposed method produces approximate posteriors that are very close to the exact distributions when those are available and also produces posteriors that cover the data-generating parameters with higher probability in the intractable case. Our results show that proper inclusion the normalising constant is crucial to the correct quantification of uncertainty and that the proposed method is an accurate and easy to implement technique to include this normalisation, being applicable to a large class of models. Key-words: Doubly-intractable; elicitation; historical data; normalisation; power prior; sensitivity analysis.
In traditional Distributional Semantic Models (DSMs) the multiple senses of a polysemous word are conflated into a single vector space representation. In this work, we propose a DSM that learns multiple distributional representations of a word based
on different topics. First, a separate DSM is trained for each topic and then each of the topic-based DSMs is aligned to a common vector space. Our unsupervised mapping approach is motivated by the hypothesis that words preserving their relative distances in different topic semantic sub-spaces constitute robust textit{semantic anchors} that define the mappings between them. Aligned cross-topic representations achieve state-of-the-art results for the task of contextual word similarity. Furthermore, evaluation on NLP downstream tasks shows that multiple topic-based embeddings outperform single-prototype models.
سجل دخول لتتمكن من نشر تعليقات
التعليقات
جاري جلب التعليقات
سجل دخول لتتمكن من متابعة معايير البحث التي قمت باختيارها
