Normalization layers are ubiquitous in modern neural networks and have long been considered essential. This work demonstrates that Transformers without normalization can achieve the same or better performance using a remarkably simple technique. We introduce Dynamic Tanh (DyT), an element-wise operation $DyT($x$) = \tanh(α$x$)$, as a drop-in replacement for normalization layers in Transformers. DyT is inspired by the observation that layer normalization in Transformers often produces tanh-like, $S$-shaped input-output mappings. By incorporating DyT, Transformers without normalization can match or exceed the performance of their normalized counterparts, mostly without hyperparameter tuning. We validate the effectiveness of Transformers with DyT across diverse settings, ranging from recognition to generation, supervised to self-supervised learning, and computer vision to language models. These findings challenge the conventional understanding that normalization layers are indispensable in modern neural networks, and offer new insights into their role in deep networks.
We propose a probabilistic interpretation of exponential dot product attention of transformers and contrastive learning based off of exponential families. The attention sublayer of transformers is equivalent to a gradient ascent step of the log normalizer, which is the log-sum-exp term in the Hopfield theory of attention. This ascent step induces a parallel expansion of points, which is counterbalanced by a contraction from layer normalization. We also state theoretical limitations of our theory and the Hopfield theory and suggest directions for resolution.
Rethink Arbitrary Style Transfer with Transformer and Contrastive Learning
College of Computer Science and Technology, Zhejiang University（浙江大学计算机科学与技术学院）；College of Computer Science and Technology, Zhejiang Gongshang University（浙江工商大学计算机科学与技术学院）
Biologically inspired spiking neural networks (SNNs) are widely used to realize ultralow-power energy consumption. However, deep SNNs are not easy to train due to the excessive firing of spiking neurons in the hidden layers. To tackle this problem, we propose a novel but simple normalization technique called postsynaptic potential normalization. This normalization removes the subtraction term from the standard normalization and uses the second raw moment instead of the variance as the division term. The spike firing can be controlled, enabling the training to proceed appropriating, by conducti…
Using an extended and formalized version of the Q/C map analysis of Poole et al. (2016), along with Neural Tangent Kernel theory, we identify the main pathologies present in deep networks that prevent them from training fast and generalizing to unseen data, and show how these can be avoided by carefully controlling the "shape" of the network's initialization-time kernel function. We then develop a method called Deep Kernel Shaping (DKS), which accomplishes this using a combination of precise parameter initialization, activation function transformations, and small architectural tweaks, all of w…
Low-resource language translation is a challenging but socially valuable NLP task. Building on recent work adapting the Transformer's normalization to this setting, we propose QKNorm, a normalization technique that modifies the attention mechanism to make the softmax function less prone to arbitrary saturation without sacrificing expressivity. Specifically, we apply $\ell_2$ normalization along the head dimension of each query and key matrix prior to multiplying them and then scale up by a learnable parameter instead of dividing by the square root of the embedding dimension. We show improvemen…
Skip connection, is a widely-used technique to improve the performance and the convergence of deep neural networks, which is believed to relieve the difficulty in optimization due to non-linearity by propagating a linear component through the neural network layers. However, from another point of view, it can also be seen as a modulating mechanism between the input and the output, with the input scaled by a pre-defined value one. In this work, we investigate how the scale factors in the effectiveness of the skip connection and reveal that a trivial adjustment of the scale will lead to spurious …
The standard normalization method for neural network (NN) models used in Natural Language Processing (NLP) is layer normalization (LN). This is different than batch normalization (BN), which is widely-adopted in Computer Vision. The preferred use of LN in NLP is principally due to the empirical observation that a (naive/vanilla) use of BN leads to significant performance degradation for NLP tasks; however, a thorough understanding of the underlying reasons for this is not always evident. In this paper, we perform a systematic study of NLP transformer models to understand why BN has a poor perf…
Chinese Academy of Sciences（中国科学院）；University of Chinese Academy of Sciences（中国科学院大学）；Peking University（北京大学）；Beijing Institute of Big Data Research（北京大数据研究院）；Microsoft Research（微软研究院）；Nankai University（南开大学）
The Transformer is widely used in natural language processing tasks. To train a Transformer however, one usually needs a carefully designed learning rate warm-up stage, which is shown to be crucial to the final performance but will slow down the optimization and bring more hyper-parameter tunings. In this paper, we first study theoretically why the learning rate warm-up stage is essential and show that the location of layer normalization matters. Specifically, we prove with mean field theory that at initialization, for the original-designed Post-LN Transformer, which places the layer normaliza…
Text normalization, or the process of transforming text into a consistent, canonical form, is crucial for speech applications such as text-to-speech synthesis (TTS). In TTS, the system must decide whether to verbalize "1995" as "nineteen ninety five" in "born in 1995" or as "one thousand nine hundred ninety five" in "page 1995". We present an experimental comparison of various Transformer-based sequence-to-sequence (seq2seq) models of text normalization for speech and evaluate them on a variety of datasets of written text aligned to its normalized spoken form. These models include variants of …
University of Notre Dame（圣母大学）；Amazon AWS AI（亚马逊AWS人工智能）
We evaluate three simple, normalization-centric changes to improve Transformer training. First, we show that pre-norm residual connections (PreNorm) and smaller initializations enable warmup-free, validation-based training with large learning rates. Second, we propose $\ell_2$ normalization with a single scale parameter (ScaleNorm) for faster training and better performance. Finally, we reaffirm the effectiveness of normalizing word embeddings to a fixed length (FixNorm). On five low-resource translation pairs from TED Talks-based corpora, these changes always converge, giving an average +1.1 …
Batch Normalization (BN) is a core and prevalent technique in accelerating the training of deep neural networks and improving the generalization on Computer Vision (CV) tasks. However, it fails to defend its position in Natural Language Processing (NLP), which is dominated by Layer Normalization (LN). In this paper, we are trying to answer why BN usually performs worse than LN in NLP tasks with Transformer models. We find that the inconsistency between training and inference of BN is the leading cause that results in the failure of BN in NLP. We define Training Inference Discrepancy (TID) to q…
Comparative Analysis of CNN and Transformer Architectures with Heart Cycle Normalization for Automated Phonocardiogram Classification
Fraunhofer Institute for Software and Systems Engineering ISST（弗劳恩霍夫软件与系统工程研究所）；Department of Computer Science, University of Applied Sciences and Arts Dortmund（应用科学大学多特蒙德计算机科学系）；Department of Computer Science, TU Dortmund University（图尔努恩大学计算机科学系）；Department of Health Informatics, School of Medicine, Witten/Herdecke University（医学学院健康信息学系，维滕/赫尔德克大学）；Institute for Medical Informatics, Biometry and Epidemiology (IMIBE), University Hospital Essen（医学信息学、生物统计与流行病学研究所（IMIBE），埃森大学医院）
Coronavirus disease or COVID-19 is an infectious disease caused by the SARS-CoV-2 virus. The first confirmed case caused by this virus was found at the end of December 2019 in Wuhan City, China. This case then spread throughout the world, including Indonesia. Therefore, the COVID-19 case was designated as a global pandemic by WHO. The growth of COVID-19 cases, especially in Indonesia, can be predicted using several approaches, such as the Deep Neural Network (DNN). One of the DNN models that can be used is Deep Transformer which can predict time series. The model is trained with several test s…
University of Udine（乌迪内大学）；University of Naples Federico II（那不勒斯费德里科二世大学）；The Hong Kong Polytechnic University（香港理工大学）；Bayer Pharmaceuticals（拜耳制药）
This paper describes the models developed by the AILAB-Udine team for the SMM4H 22 Shared Task. We explored the limits of Transformer based models on text classification, entity extraction and entity normalization, tackling Tasks 1, 2, 5, 6 and 10. The main take-aways we got from participating in different tasks are: the overwhelming positive effects of combining different architectures when using ensemble learning, and the great potential of generative models for term normalization.
Norwegian University of Science and Technology（挪威科技大学）
Time series forecasting is an important problem, with many real world applications. Ensembles of deep neural networks have recently achieved impressive forecasting accuracy, but such large ensembles are impractical in many real world settings. Transformer models been successfully applied to a diverse set of challenging problems. We propose a novel adaptation of the original Transformer architecture focusing on the task of time series forecasting, called Persistence Initialization. The model is initialized as a naive persistence model by using a multiplicative gating mechanism combined with a r…
Learning in Compact Spaces with Approximately Normalized Transformer
University of Freiburg（弗赖堡大学）；ELLIS Institute Tübingen（图宾根ELLIS研究所）；Open-Sci Collective（开放科学集体）；LAION；Jülich Supercomputing Centre (JSC)（尤利奇超级计算中心）；Prior Labs（Prior实验室）
Zhejiang University（浙江大学）；Microsoft Research Asia（微软亚洲研究院）
The Transformer architecture consists of self-attention and feed-forward networks (FFNs) which can be viewed as key-value memories according to previous works. However, FFN and traditional memory utilize different activation functions (i.e., ReLU and Softmax respectively), which makes them not equivalent. In this paper, we first rebuild the connections between FFN and key-value memory by conducting extensive studies on ReLU and Softmax, and find they are equivalent when adding an additional layer normalization module on Softmax. In addition, ReLU outperforms Softmax on both FFN and key-value m…
Exploiting Layer Normalization Fine-tuning in Visual Transformer Foundation Models for Classification
Shanghai Academy of Artificial Intelligence for Science（上海人工智能科学研究院）；Xi’an Jiaotong-Liverpool University（西安交通大学利物浦大学）；University of Liverpool（利物浦大学）；AI 3 Fudan University（复旦大学）；Duke Kunshan University（杜克-昆山大学）；BII, A ∗ STAR；Zhejiang University（浙江大学）
School of Engineering（工程学院）；Electrical Engineering and Computer Engineering Department（电子工程与计算机工程系）；Quebec University at Trois-Rivières（魁北克大学 Trois-Rivières）
Normalization is a key point in transformer training . In Dynamic Tanh (DyT), the author demonstrated that Tanh can be used as an alternative layer normalization (LN) and confirmed the effectiveness of the idea. But Tanh itself faces orthogonality, linearity and distortion problems. Due to that, his proposition cannot be reliable. So we propose a Holonorm (hn) which has residual connections and nonlinearity. Holonorm is suitable for replacing Tanh in the context of normalization. Although the HoloNorm expression could be similar to the softsign function in dimension one, softsign is a componen…
Normalization layers are a staple in state-of-the-art deep neural network architectures. They are widely believed to stabilize training, enable higher learning rate, accelerate convergence and improve generalization, though the reason for their effectiveness is still an active research topic. In this work, we challenge the commonly-held beliefs by showing that none of the perceived benefits is unique to normalization. Specifically, we propose fixed-update initialization (Fixup), an initialization motivated by solving the exploding and vanishing gradient problem at the beginning of training via…
School of Computer Science and Engineering, Nanyang Technological University（南洋理工大学计算机科学与工程学院）
We propose a generalization of transformer neural network architecture for arbitrary graphs. The original transformer was designed for Natural Language Processing (NLP), which operates on fully connected graphs representing all connections between the words in a sequence. Such architecture does not leverage the graph connectivity inductive bias, and can perform poorly when the graph topology is important and has not been encoded into the node features. We introduce a graph transformer with four new properties compared to the standard model. First, the attention mechanism is a function of the n…
In this paper, we apply the self-attention from the state-of-the-art Transformer in Attention Is All You Need for the first time to a data-driven operator learning problem related to partial differential equations. An effort is put together to explain the heuristics of, and to improve the efficacy of the attention mechanism. By employing the operator approximation theory in Hilbert spaces, it is demonstrated for the first time that the softmax normalization in the scaled dot-product attention is sufficient but not necessary. Without softmax, the approximation capacity of a linearized Transform…
Batch normalization is a key component of most image classification models, but it has many undesirable properties stemming from its dependence on the batch size and interactions between examples. Although recent work has succeeded in training deep ResNets without normalization layers, these models do not match the test accuracies of the best batch-normalized networks, and are often unstable for large learning rates or strong data augmentations. In this work, we develop an adaptive gradient clipping technique which overcomes these instabilities, and design a significantly improved class of Nor…
