Language Modeling

CTAL is a pre-training framework for strong audio-and-language representations with a Transformer, which aims to learn the intra-modality and inter-modalities connections between audio and language through two proxy tasks on a large amount of audio- and-language pairs: masked language modeling and masked cross-modal acoustic modeling. The pre-trained model is a Transformer for Audio and Language, i.e., CTAL, which consists of two modules, a language stream encoding module which adapts word as input element, and a text-referred audio stream encoder module which accepts both frame-level Mel-spectrograms and token-level output embeddings from the language stream

DynaBERT is a BERT-variant which can flexibly adjust the size and latency by selecting adaptive width and depth. The training process of DynaBERT includes first training a width-adaptive BERT and then allowing both adaptive width and depth, by distilling knowledge from the full-sized model to small sub-networks. Network rewiring is also used to keep the more important attention heads and neurons shared by more sub-networks. A two-stage procedure is used to train DynaBERT. First, using knowledge distillation (dashed lines) to transfer the knowledge from a fixed teacher model to student sub-networks with adaptive width in DynaBERTW. Then, using knowledge distillation (dashed lines) to transfer the knowledge from a trained DynaBERTW to student sub-networks with adaptive width and depth in DynaBERT.

An End-to-End Memory Network is a neural network with a recurrent attention model over a possibly large external memory. The architecture is a form of Memory Network, but unlike the model in that work, it is trained end-to-end, and hence requires significantly less supervision during training. It can also be seen as an extension of RNNsearch to the case where multiple computational steps (hops) are performed per output symbol. The model takes a discrete set of inputs $x\_{1}, \dots, x\_{n}$ that are to be stored in the memory, a query $q$, and outputs an answer $a$. Each of the $x\_{i}$, $q$, and $a$ contains symbols coming from a dictionary with $V$ words. The model writes all $x$ to the memory up to a fixed buffer size, and then finds a continuous representation for the $x$ and $q$. The continuous representation is then processed via multiple hops to output $a$.

A Feedback Transformer is a type of sequential transformer that exposes all previous representations to all future representations, meaning the lowest representation of the current timestep is formed from the highest-level abstract representation of the past. This feedback nature allows this architecture to perform recursive computation, building stronger representations iteratively upon previous states. To achieve this, the self-attention mechanism of the standard Transformer is modified so it attends to higher level representations rather than lower ones.

Fraternal Dropout is a regularization method for recurrent neural networks that trains two identical copies of an RNN (that share parameters) with different dropout masks while minimizing the difference between their (pre-softmax) predictions. This encourages the representations of RNNs to be invariant to dropout mask, thus being robust.

MacBERT is a Transformer-based model for Chinese NLP that alters RoBERTa in several ways, including a modified masking strategy. Instead of masking with [MASK] token, which never appears in the fine-tuning stage, MacBERT masks the word with its similar word. Specifically MacBERT shares the same pre-training tasks as BERT with several modifications. For the MLM task, the following modifications are performed: - Whole word masking is used as well as Ngram masking strategies for selecting candidate tokens for masking, with a percentage of 40%, 30%, 20%, 10% for word-level unigram to 4-gram. - Instead of masking with [MASK] token, which never appears in the fine-tuning stage, similar words are used for the masking purpose. A similar word is obtained by using Synonyms toolkit which is based on word2vec similarity calculations. If an N-gram is selected to mask, we will find similar words individually. In rare cases, when there is no similar word, we will degrade to use random word replacement. - A percentage of 15% input words is used for masking, where 80% will replace with similar words, 10% replace with a random word, and keep with original words for the rest of 10%.

MaxUp is an adversarial data augmentation technique for improving the generalization performance of machine learning models. The idea is to generate a set of augmented data with some random perturbations or transforms, and minimize the maximum, or worst case loss over the augmented data. By doing so, we implicitly introduce a smoothness or robustness regularization against the random perturbations, and hence improve the generation performance. For example, in the case of Gaussian perturbation, MaxUp is asymptotically equivalent to using the gradient norm of the loss as a penalty to encourage smoothness.

Mesh-TensorFlow is a language for specifying a general class of distributed tensor computations. Where data-parallelism can be viewed as splitting tensors and operations along the "batch" dimension, in Mesh-TensorFlow, the user can specify any tensor dimensions to be split across any dimensions of a multi-dimensional mesh of processors. A MeshTensorFlow graph compiles into a SPMD program consisting of parallel operations coupled with collective communication primitives such as Allreduce.

SHA-RNN, or Single Headed Attention RNN, is a recurrent neural network, and language model when combined with an embedding input and softmax classifier, based on a core LSTM component and a single-headed attention module. Other design choices include a Boom feedforward layer and the use of layer normalization. The guiding principles of the author were to ensure simplicity in the architecture and to keep computational costs bounded (the model was originally trained with a single GPU).

SRU++ is a self-attentive recurrent unit that combines fast recurrence and attention for sequence modeling, extending the SRU unit. The key modification of SRU++ is to incorporate more expressive non-linear operations into the recurrent network. Specifically, given the input sequence represented as a matrix $\mathbf{X} \in \mathbb{R}^{L \times d}$, the attention component computes the query, key and value representations using the following multiplications, $$ \mathbf{Q} =\mathbf{W}^{q} \mathbf{X}^{\top} $$ $$ \mathbf{K} =\mathbf{W}^{k} \mathbf{Q} \\ $$ $$ \mathbf{V} =\mathbf{W}^{v} \mathbf{Q} $$ where $\mathbf{W}^{q} \in \mathbb{R}^{d^{\prime} \times d}, \mathbf{W}^{k}, \mathbf{W}^{v} \in \mathbb{R}^{d^{\prime} \times d^{\prime}}$ are model parameters. $d^{\prime}$ is the attention dimension that is typically much smaller than $d$. Note that the keys $\mathbf{K}$ and values $\mathbf{V}$ are computed using $\mathbf{Q}$ instead of $\mathbf{X}$ such that the weight matrices $\mathbf{W}^{k}$ and $\mathbf{W}^{v}$ are significantly smaller. Next, we compute a weighted average output $\mathbf{A} \in \mathbb{R}^{d^{\prime} \times L}$ using scaled dot-product attention: $$ \mathbf{A}^{\top}=\operatorname{softmax}\left(\frac{\mathbf{Q}^{\top} \mathbf{K}}{\sqrt{d^{\prime}}}\right) \mathbf{V}^{\top} $$ The final output $U$ required by the elementwise recurrence is obtained by another linear projection, $$ \mathbf{U}^{\top}=\mathbf{W}^{o}(\mathbf{Q}+\alpha \cdot \mathbf{A}) $$ where $\alpha \in \mathbb{R}$ is a learned scalar and $\mathbf{W}\_{o} \in \mathbb{R}^{3 d \times d^{\prime}}$ is a parameter matrix. $\mathbf{Q}+\alpha \cdot \mathbf{A}$ is a residual connection which improves gradient propagation and stabilizes training. We initialize $\alpha$ to zero and as a result, $$ \mathbf{U}^{\top}=\mathbf{W}^{o} \mathbf{Q}=\left(\mathbf{W}^{o} \mathbf{W}^{q}\right) \mathbf{X}^{\top} $$ initially falls back to a linear transformation of the input $X$ skipping the attention transformation. Intuitively, skipping attention encourages leveraging recurrence to capture sequential patterns during early stage of training. As $|\alpha|$ grows, the attention mechanism can learn long-range dependencies for the model. In addition, $\mathbf{W}^{o} \mathbf{W}^{q}$ can be interpreted as applying a matrix factorization trick with a small inner dimension $d^{\prime}<d$, reducing the total number of parameters. The Figure compares the differences of SRU, SRU with this factorization trick (but without attention), and SRU++. The last modification is adding layer normalization to each SRU++ layer. We apply normalization after the attention operation and before the matrix multiplication with $\mathbf{W}^{o}$ $$ \mathbf{U}^{\top}=\mathbf{W}^{o} \operatorname{layernorm}(\mathbf{Q}+\alpha \cdot \mathbf{A}) $$ This implementation is post-layer normalization in which the normalization is added after the residual connection.

A Sandwich Transformer is a variant of a Transformer that reorders sublayers in the architecture to achieve better performance. The reordering is based on the authors' analysis that models with more self-attention toward the bottom and more feedforward sublayers toward the top tend to perform better in general.

Please enter a description about the method here

Sparse Sinkhorn Attention is an attention mechanism that reduces the memory complexity of the dot-product attention mechanism and is capable of learning sparse attention outputs. It is based on the idea of differentiable sorting of internal representations within the self-attention module. SSA incorporates a meta sorting network that learns to rearrange and sort input sequences. Sinkhorn normalization is used to normalize the rows and columns of the sorting matrix. The actual SSA attention mechanism then acts on the block sorted sequences.

The BP-Transformer (BPT) is a type of Transformer that is motivated by the need to find a better balance between capability and computational complexity for self-attention. The architecture partitions the input sequence into different multi-scale spans via binary partitioning (BP). It incorporates an inductive bias of attending the context information from fine-grain to coarse-grain as the relative distance increases. The farther the context information is, the coarser its representation is. BPT can be regard as graph neural network, whose nodes are the multi-scale spans. A token node can attend the smaller-scale span for the closer context and the larger-scale span for the longer distance context. The representations of nodes are updated with Graph Self-Attention.

Child-Tuning is a fine-tuning technique that updates a subset of parameters (called child network) of large pretrained models via strategically masking out the gradients of the non-child network during the backward process. It decreases the hypothesis space of the model via a task-specific mask applied to the full gradients, helping to effectively adapt the large-scale pretrained model to various tasks and meanwhile aiming to maintain its original generalization ability.

CPT, or Chinese Pre-trained Unbalanced Transformer, is a pre-trained unbalanced Transformer for Chinese natural language understanding (NLU) and natural language generation (NLG) tasks. CPT consists of three parts: a shared encoder, an understanding decoder, and a generation decoder. Two specific decoders with a shared encoder are pre-trained with masked language modeling (MLM) and denoising auto-encoding (DAE) tasks, respectively. With the partially shared architecture and multi-task pre-training, CPT can (1) learn specific knowledge of both NLU or NLG tasks with two decoders and (2) be fine-tuned flexibly that fully exploits the potential of the model. Two specific decoders with a shared encoder are pre-trained with masked language modeling (MLM) and denoising auto-encoding (DAE) tasks, respectively. With the partially shared architecture and multi-task pre-training, CPT can (1) learn specific knowledge of both NLU or NLG tasks with two decoders and (2) be fine-tuned flexibly that fully exploits the potential of the model.

DExTra, or Deep and Light-weight Expand-reduce Transformation, is a light-weight expand-reduce transformation that enables learning wider representations efficiently. DExTra maps a $d\_{m}$ dimensional input vector into a high dimensional space (expansion) and then reduces it down to a $d\_{o}$ dimensional output vector (reduction) using $N$ layers of group transformations. During these expansion and reduction phases, DExTra uses group linear transformations because they learn local representations by deriving the output from a specific part of the input and are more efficient than linear transformations. To learn global representations, DExTra shares information between different groups in the group linear transformation using feature shuffling Formally, the DExTra transformation is controlled by five configuration parameters: (1) depth $N$, (2) width multiplier $m\_{w}$, (3) input dimension $d\_{m}$, (4) output dimension $d\_{o}$, and (5) maximum groups $g\_{max}$ in a group linear transformation. In the expansion phase, DExTra projects the $d\_{m}$-dimensional input to a high-dimensional space, $d\_{max} = m\_{w}d\_{m}$, linearly using $\text{ceil}\left(\frac{N}{2}\right)$ layers. In the reduction phase, DExTra projects the $d\_{max}$-dimensional vector to a $d\_{o}$-dimensional space using the remaining $N -\text{ceil}\left(\frac{N}{2}\right)$ layers. Mathematically, we define the output $Y$ at each layer $l$ as: $$ \mathbf{Y}\_{l} = \mathcal{F}\left(\mathbf{X}, \mathbf{W}^{l}, \mathbf{b}^{l}, g^{l}\right) \text{ if } l=1 $$ $$ \mathbf{Y}\_{l} = \mathcal{F}\left(\mathcal{H}\left(\mathbf{X}, \mathbf{Y}^{l-1}\right), \mathbf{W}^{l}, \mathbf{b}^{l}, g^{l}\right) \text{ Otherwise } $$ where the number of groups at each layer $l$ are computed as: $$ g^{l} = \text{min}\left(2^{l-1}, g\_{max}\right), 1 \leq l \leq \text{ceil}\left(N/2\right) $$ $$ g^{N-l}, \text{Otherwise}$$ In the above equations, $\mathcal{F}$ is a group linear transformation function. The function $\mathcal{F}$ takes the input $\left(\mathbf{X} \text{ or } \mathcal{H}\left(\mathbf{X}, \mathbf{Y}^{l-1}\right) \right)$, splits it into $g^{l}$ groups, and then applies a linear transformation with learnable parameters $\mathbf{W}^{l}$ and bias $\mathbf{b}^{l}$ to each group independently. The outputs of each group are then concatenated to produce the final output $\mathbf{Y}^{l}$. The function $\mathcal{H}$ first shuffles the output of each group in $\mathbf{Y}^{l−1}$ and then combines it with the input $\mathbf{X}$ using an input mixer connection. In the authors' experiments, they use $g\_{max} = \text{ceil}\left(\frac{d\_{m}}{32}\right)$ so that each group has at least 32 input elements. Note that (i) group linear transformations reduce to linear transformations when $g^{l} = 1$, and (ii) DExTra is equivalent to a multi-layer perceptron when $g\_{max} = 1$.

DeLiGHT is a transformer architecture that delivers parameter efficiency improvements by (1) within each Transformer block using DExTra, a deep and light-weight transformation, allowing for the use of single-headed attention and bottleneck FFN layers and (2) across blocks using block-wise scaling, that allows for shallower and narrower DeLighT blocks near the input and wider and deeper DeLighT blocks near the output.

A DeLighT Block is a block used in the DeLighT transformer architecture. It uses a DExTra transformation to reduce the dimensionality of the vectors entered into the attention layer, where a single-headed attention module is used. Since the DeLighT block learns wider representations of the input across different layers using DExTra, it enables the authors to replace multi-head attention with single-head attention. This is then followed by a light-weight FFN which, rather than expanding the dimension (as in normal Transformers which widen to a dimension 4x the size), imposes a bottleneck and squeezes the dimensions. Again, the reason for this is that the DExTra transformation has already incorporated wider representations so we can squeeze instead at this layer.

ESPNetv2 is a convolutional neural network that utilises group point-wise and depth-wise dilated separable convolutions to learn representations from a large effective receptive field with fewer FLOPs and parameters.

Factorized Dense Synthesized Attention is a synthesized attention mechanism, similar to dense synthesized attention, but we factorize the outputs to reduce parameters and prevent overfitting. It was proposed as part of the Synthesizer architecture. The factorized variant of the dense synthesizer can be expressed as follows: $$A, B = F\_{A}\left(X\_{i}\right), F\_{B}\left(X\_{i}\right)$$ where $F\_{A}\left(.\right)$ projects input $X\_{i}$ into $a$ dimensions, $F\_B\left(.\right)$ projects $X\_{i}$ to $b$ dimensions, and $a \text{ x } b = l$. The output of the factorized module is now written as: $$ Y = \text{Softmax}\left(C\right)G\left(X\right) $$ where $C = H\_{A}\left(A\right) * H\_{B}\left(B\right)$, where $H\_{A}$, $H\_{B}$ are tiling functions and $C \in \mathbb{R}^{l \text{ x } l}$. The tiling function simply duplicates the vector $k$ times, i.e., $\mathbb{R}^{l} \rightarrow \mathbb{R}^{lk}$. In this case, $H\_{A}\left(\right)$ is a projection of $\mathbb{R}^{a} \rightarrow \mathbb{R}^{ab}$ and $H\_{B}\left(\right)$ is a projection of $\mathbb{R}^{b} \rightarrow \mathbb{R}^{ba}$. To avoid having similar values within the same block, we compose the outputs of $H\_{A}$ and $H\_{B}$.

Factorized Random Synthesized Attention, introduced with the Synthesizer architecture, is similar to factorized dense synthesized attention but for random synthesizers. Letting $R$ being a randomly initialized matrix, we factorize $R$ into low rank matrices $R\_{1}, R\_{2} \in \mathbb{R}^{l\text{ x}k}$ in the attention function: $$ Y = \text{Softmax}\left(R\_{1}R\_{2}^{T}\right)G\left(X\right) . $$ Here $G\left(.\right)$ is a parameterized function that is equivalent to $V$ in Scaled Dot-Product Attention. For each head, the factorization reduces the parameter costs from $l^{2}$ to $2\left(lk\right)$ where $k << l$ and hence helps prevent overfitting. In practice, we use a small value of $k = 8$. The basic idea of a Random Synthesizer is to not rely on pairwise token interactions or any information from individual token but rather to learn a task-specific alignment that works well globally across many samples.

K3M is a multi-modal pretraining method for e-commerce product data that introduces knowledge modality to correct the noise and supplement the missing of image and text modalities. The modal-encoding layer extracts the features of each modality. The modal-interaction layer is capable of effectively modeling the interaction of multiple modalities, where an initial-interactive feature fusion model is designed to maintain the independence of image modality and text modality, and a structure aggregation module is designed to fuse the information of image, text, and knowledge modalities. K3M is pre-trained with three pretraining tasks, including masked object modeling (MOM), masked language modeling (MLM), and link prediction modeling (LPM).

NormFormer is a type of Pre-LN transformer that adds three normalization operations to each layer: a Layer Norm after self attention, head-wise scaling of self-attention outputs, and a Layer Norm after the first fully connected layer. The modifications introduce a small number of additional learnable parameters, which provide a cost-effective way for each layer to change the magnitude of its features, and therefore the magnitude of the gradients to subsequent components.

PermuteFormer is a Performer-based model with relative position encoding that scales linearly on long sequences. PermuteFormer applies position-dependent transformation on queries and keys to encode positional information into the attention module. This transformation is carefully crafted so that the final output of self-attention is not affected by absolute positions of tokens. Each token’s query / key feature is illustrated as a row of blocks in the figure, and its elements are marked with different colors. The position-aware permutation permutes elements of each token’s query / key feature along the head size dimension in each attention head. Depending on the token’s position, the permutation applied to query / key feature is different.

The Pointer Sentinel-LSTM mixture model is a type of recurrent neural network that combines the advantages of standard softmax classifiers with those of a pointer component for effective and efficient language modeling. Rather than relying on the RNN hidden state to decide when to use the pointer, the model allows the pointer component itself to decide when to use the softmax vocabulary through a sentinel.

RealFormer is a type of Transformer based on the idea of residual attention. It adds skip edges to the backbone Transformer to create multiple direct paths, one for each type of attention module. It adds no parameters or hyper-parameters. Specifically, RealFormer uses a Post-LN style Transformer as backbone and adds skip edges to connect Multi-Head Attention modules in adjacent layers.

SMITH, or Siamese Multi-depth Transformer-based Hierarchical Encoder, is a Transformer-based model for document representation learning and matching. It contains several design choices to adapt self-attention models for long text inputs. For the model pre-training, a masked sentence block language modeling task is used in addition to the original masked word language model task used in BERT, to capture sentence block relations within a document. Given a sequence of sentence block representation, the document level Transformers learn the contextual representation for each sentence block and the final document representation.

The Sinkhorn Transformer is a type of transformer that uses Sparse Sinkhorn Attention as a building block. This component is a plug-in replacement for dense fully-connected attention (as well as local attention, and sparse attention alternatives), and allows for reduced memory complexity as well as sparse attention.

SortCut Sinkhorn Attention is a variant of Sparse Sinkhorn Attention where a post-sorting truncation of the input sequence is performed, essentially performing a hard top-k operation on the input sequence blocks within the computational graph. While most attention models mainly re-weight or assign near-zero weights during training, this allows for explicitly and dynamically truncate the input sequence. Specifically: $$ Y = \text{Softmax}\left(Q{\psi\_{S}}\left(K\right)^{T}\_{\left[:n\right]}\right)\psi\_{S}\left(V\right)\_{\left[:n\right]} $$ where $n$ is the Sortfut budget hyperparameter.

A Strided EESP unit is based on the EESP Unit but is modified to learn representations more efficiently at multiple scales. Depth-wise dilated convolutions are given strides, an average pooling operation is added instead of an identity connection, and the element-wise addition operation is replaced with a concatenation operation, which helps in expanding the dimensions of feature maps efficiently.

A Time-aware Large Kernel (TaLK) convolution is a type of temporal convolution that learns the kernel size of a summation kernel for each time-step instead of learning the kernel weights as in a typical convolution operation. For each time-step, a function is responsible for predicting the appropriate size of neighbor representations to use in the form of left and right offsets relative to the time-step.

Talking-Heads Attention is a variation on multi-head attention which includes linear projections across the attention-heads dimension, immediately before and after the softmax operation. In multi-head attention, the different attention heads perform separate computations, which are then summed at the end. Talking-Heads Attention breaks that separation. Two additional learned linear projections are inserted, $P\_{l}$ and $P\_{w}$, which transform the attention-logits and the attention weights respectively, moving information across attention heads. Instead of one "heads" dimension $h$ across the whole computation, we now have three separate heads dimensions: $h\_{k}$, $h$, and $h\_{v}$, which can optionally differ in size (number of "heads"). $h\_{k}$ refers to the number of attention heads for the keys and the queries. $h$ refers to the number of attention heads for the logits and the weights, and $h\_{v}$ refers to the number of attention heads for the values.

Unigram Segmentation is a subword segmentation algorithm based on a unigram language model. It provides multiple segmentations with probabilities. The language model allows for emulating the noise generated during the segmentation of actual data. The unigram language model makes an assumption that each subword occurs independently, and consequently, the probability of a subword sequence $\mathbf{x} = (x_1,\ldots,x_M)$ is formulated as the product of the subword occurrence probabilities $p(x_i)$: $$ P(\mathbf{x}) = \prod_{i=1}^{M} p(x_i), \\\\ \forall i\,\, x_i \in \mathcal{V},\,\,\, \sum_{x \in \mathcal{V}} p(x) = 1, \nonumber $$ where $\mathcal{V}$ is a pre-determined vocabulary. The most probable segmentation $\mathbf{x}^*$ for the input sentence $X$ is then given by: $$ \mathbf{x}^{*} = \text{argmax}_{\mathbf{x} \in \mathcal{S}(X)} P(\mathbf{x}), $$ where $\mathcal{S}(X)$ is a set of segmentation candidates built from the input sentence $X$. $\mathbf{x}^*$ is obtained with the Viterbi algorithm.

XGPT is a method of cross-modal generative pre-training for image captioning designed to pre-train text-to-image caption generators through three novel generation tasks, including image-conditioned masked language modeling (IMLM), image-conditioned denoising autoencoding (IDA), and text-conditioned image feature generation (TIGF). The pre-trained XGPT can be fine-tuned without any task-specific architecture modifications and build strong image captioning models.