Language Modelling

Residual Shuffle-Exchange Network is an efficient alternative to models using an attention mechanism that allows the modelling of long-range dependencies in sequences in O(n log n) time. This model achieved state-of-the-art performance on the MusicNet dataset for music transcription while being able to run inference on a single GPU fast enough to be suitable for real-time audio processing.

The Beneš block is a computation-efficient alternative to dense attention, enabling the modelling of long-range dependencies in O(n log n) time. In comparison, dense attention which is commonly used in Transformers has O(n^2) complexity. In music, dependencies occur on several scales, including on a coarse scale which requires processing very long sequences. Beneš blocks have been used in Residual Shuffle-Exchange Networks to achieve state-of-the-art results in music transcription. Beneš blocks have a ‘receptive field’ of the size of the whole sequence, and it has no bottleneck. These properties hold for dense attention but have not been shown for many sparse attention and dilated convolutional architectures.

Dynamic Sparse Training method based on using the gradient norm to select which weights to remove/add in each mask update.

VL-T5 is a unified framework that learns different tasks in a single architecture with the same language modeling objective, i.e., multimodal conditional text generation. The model learns to generate labels in text based on the visual and textual inputs. In contrast to other existing methods, the framework unifies tasks as generating text labels conditioned on multimodal inputs. This allows the model to tackle vision-and-language tasks with unified text generation objective. The models use text prefixes to adapt to different tasks.

The composed video retrieval (CoVR) task is a new task, where the goal is to find a video that matches both a query image and a query text. The query image represents a visual concept that the user is interested in, and the query text specifies how the concept should be modified or refined. For example, given an image of a fountain and the text during show at night, the CoVR task is to retrieve a video that shows the fountain at night with a show.

Feedback Memory is a type of attention module used in the Feedback Transformer architecture. It allows a transformer to to use the most abstract representations from the past directly as inputs for the current timestep. This means that the model does not form its representation in parallel, but sequentially token by token. More precisely, we replace the context inputs to attention modules with memory vectors that are computed over the past, i.e.: $$ \mathbf{z}^{l}\_{t} = \text{Attn}\left(\mathbf{x}^{l}\_{t}, \left[\mathbf{m}\_{t−\tau}, \dots, \mathbf{m}\_{t−1}\right]\right) $$ where a memory vector $\mathbf{m}\_{t}$ is computed by summing the representations of each layer at the $t$-th time step: $$ \mathbf{m}\_{t} = \sum^{L}\_{l=0}\text{Softmax}\left(w^{l}\right)\mathbf{x}\_{t}^{l} $$ where $w^{l}$ are learnable scalar parameters. Here $l = 0$ corresponds to token embeddings. The weighting of different layers by a softmax output gives the model more flexibility as it can average them or select one of them. This modification of the self-attention input adapts the computation of the Transformer from parallel to sequential, summarized in the Figure. Indeed, it gives the ability to formulate the representation $\mathbf{x}^{l}\_{t+1}$ based on past representations from any layer $l'$, while in a standard Transformer this is only true for $l > l'$. This change can be viewed as exposing all previous computations to all future computations, providing better representations of the input. Such capacity would allow much shallower models to capture the same level of abstraction as a deeper architecture.

Method introduces a novel unlabeled debiasing technique which works on classification task to reduce the bias of the transformer based language models on downstream classification task. In their method authors use the classes as metric for regularization and punish the network if the embedding produced by the model are far from each other. by doing so the authors claim to be able to reduce the domain shift caused by any unwanted attribute information hence results in fair embedding.

LayerDrop is a form of structured dropout for Transformer models which has a regularization effect during training and allows for efficient pruning at inference time. It randomly drops layers from the Transformer according to an "every other" strategy where pruning with a rate $p$ means dropping the layers at depth $d$ such that $d = 0\left\(\text{mod}\left(\text{floor}\left(\frac{1}{p}\right)\right)\right)$.

A Neural Cache, or a Continuous Cache, is a module for language modelling which stores previous hidden states in memory cells. They are then used as keys to retrieve their corresponding word, that is the next word. There is no transformation applied to the storage during writing and reading. More formally it exploits the hidden representations $h\_{t}$ to define a probability distribution over the words in the cache. As illustrated in the Figure, the cache stores pairs $\left(h\_{i}, x\_{i+1}\right)$ of a hidden representation, and the word which was generated based on this representation (the vector $h\_{i}$ encodes the history $x\_{i}, \dots, x\_{1}$). At time $t$, we then define a probability distribution over words stored in the cache based on the stored hidden representations and the current one $h\_{t}$ as: $$ p\_{cache}\left(w | h\_{1\dots{t}}, x\_{1\dots{t}}\right) \propto \sum^{t-1}\_{i=1}\mathcal{1}\_{\text{set}\left(w=x\_{i+1}\right)} \exp\left(θ\_{h}>h\_{t}^{T}h\_{i}\right) $$ where the scalar $\theta$ is a parameter which controls the flatness of the distribution. When $\theta$ is equal to zero, the probability distribution over the history is uniform, and the model is equivalent to a unigram cache model.

VL-BERT is pre-trained on a large-scale image-captions dataset together with text-only corpus. The input to the model are either words from the input sentences or regions-of-interest (RoI) from input images. It can be fine-tuned to fit most visual-linguistic downstream tasks. Its backbone is a multi-layer bidirectional Transformer encoder, modified to accommodate visual contents, and new type of visual feature embedding to the input feature embeddings. VL-BERT takes both visual and linguistic elements as input, represented as RoIs in images and subwords in input sentences. Four different types of embeddings are used to represent each input: token embedding, visual feature embedding, segment embedding, and sequence position embedding. VL-BERT is pre-trained using Conceptual Captions and text-only datasets. Two pre-training tasks are used: masked language modeling with visual clues, and masked RoI classification with linguistic clues.

Virtual Data Augmentation, or VDA, is a framework for robustly fine-tuning pre-trained language model. Based on the original token embeddings, a multinomial mixture for augmenting virtual data is constructed, where a masked language model guarantees the semantic relevance and the Gaussian noise provides the augmentation diversity. Furthermore, a regularized training strategy is proposed to balance the two aspects.

The Compressive Transformer is an extension to the Transformer which maps past hidden activations (memories) to a smaller set of compressed representations (compressed memories). The Compressive Transformer uses the same attention mechanism over its set of memories and compressed memories, learning to query both its short-term granular memory and longer-term coarse memory. It builds on the ideas of Transformer-XL which maintains a memory of past activations at each layer to preserve a longer history of context. The Transformer-XL discards past activations when they become sufficiently old (controlled by the size of the memory). The key principle of the Compressive Transformer is to compress these old memories, instead of discarding them, and store them in an additional compressed memory. At each time step $t$, we discard the oldest compressed memories (FIFO) and then the oldest $n$ states from ordinary memory are compressed and shifted to the new slot in compressed memory. During training, the compressive memory component is optimized separately from the main language model (separate training loop).

ControlVAE is a variational autoencoder (VAE) framework that combines the automatic control theory with the basic VAE to stabilize the KL-divergence of VAE models to a specified value. It leverages a non-linear PI controller, a variant of the proportional-integral-derivative (PID) control, to dynamically tune the weight of the KL-divergence term in the evidence lower bound (ELBO) using the output KL-divergence as feedback. This allows for control of the KL-divergence to a desired value (set point), which is effective in avoiding posterior collapse and learning disentangled representations.

An EESP Unit, or Extremely Efficient Spatial Pyramid of Depth-wise Dilated Separable Convolutions, is an image model block designed for edge devices. It was proposed as part of the ESPNetv2 CNN architecture. This building block is based on a reduce-split-transform-merge strategy. The EESP unit first projects the high-dimensional input feature maps into low-dimensional space using groupwise pointwise convolutions and then learns the representations in parallel using depthwise dilated separable convolutions with different dilation rates. Different dilation rates in each branch allow the EESP unit to learn the representations from a large effective receptive field. To remove the gridding artifacts caused by dilated convolutions, the EESP fuses the feature maps using hierarchical feature fusion (HFF).

Gated Transformer-XL, or GTrXL, is a Transformer-based architecture for reinforcement learning. It introduces architectural modifications that improve the stability and learning speed of the original Transformer and XL variant. Changes include: - Placing the layer normalization on only the input stream of the submodules. A key benefit to this reordering is that it now enables an identity map from the input of the transformer at the first layer to the output of the transformer after the last layer. This is in contrast to the canonical transformer, where there are a series of layer normalization operations that non-linearly transform the state encoding. - Replacing residual connections with gating layers. The authors' experiments found that GRUs were the most effective form of gating.

A Gated Convolutional Network is a type of language model that combines convolutional networks with a gating mechanism. Zero padding is used to ensure future context can not be seen. Gated convolutional layers can be stacked on top of other hierarchically. Model predictions are then obtained with an adaptive softmax layer.

OPT-IML is a version of OPT fine-tuned on a large collection of 1500+ NLP tasks divided into various task categories.

Online Normalization is a normalization technique for training deep neural networks. To define Online Normalization. we replace arithmetic averages over the full dataset in with exponentially decaying averages of online samples. The decay factors $\alpha\_{f}$ and $\alpha\_{b}$ for forward and backward passes respectively are hyperparameters for the technique. We allow incoming samples $x\_{t}$, such as images, to have multiple scalar components and denote feature-wide mean and variance by $\mu\left(x\_{t}\right)$ and $\sigma^{2}\left(x\_{t}\right)$. The algorithm also applies to outputs of fully connected layers with only one scalar output per feature. In fact, this case simplifies to $\mu\left(x\_{t}\right) = x\_{t}$ and $\sigma\left(x\_{t}\right) = 0$. Denote scalars $\mu\_{t}$ and $\sigma\_{t}$ to denote running estimates of mean and variance across all samples. The subscript $t$ denotes time steps corresponding to processing new incoming samples. Online Normalization uses an ongoing process during the forward pass to estimate activation means and variances. It implements the standard online computation of mean and variance generalized to processing multi-value samples and exponential averaging of sample statistics. The resulting estimates directly lead to an affine normalization transform. $$ y\_{t} = \frac{x\_{t} - \mu\_{t-1}}{\sigma\_{t-1}} $$ $$ \mu\_{t} = \alpha\_{f}\mu\_{t-1} + \left(1-\alpha\_{f}\right)\mu\left(x\_{t}\right) $$ $$ \sigma^{2}\_{t} = \alpha\_{f}\sigma^{2}\_{t-1} + \left(1-\alpha\_{f}\right)\sigma^{2}\left(x\_{t}\right) + \alpha\_{f}\left(1-\alpha\_{f}\right)\left(\mu\left(x\_{t}\right) - \mu\_{t-1}\right)^{2} $$

Probabilistically Masked Language Model, or PMLM, is a type of language model that utilizes a probabilistic masking scheme, aiming to bridge the gap between masked and autoregressive language models. The basic idea behind the connection of two categories of models is similar to MADE by Germain et al (2015). PMLM is a masked language model with a probabilistic masking scheme, which defines the way sequences are masked by following a probabilistic distribution. The authors employ a simple uniform distribution of the masking ratio and name the model as u-PMLM.

• Parallel Layers – We use a “parallel” formulation in each Transformer block (Wang & Komatsuzaki, 2021), rather than the standard “serialized” formulation. Specifically, the standard formulation can be written as: y = x + MLP(LayerNorm(x + Attention(LayerNorm(x))) Whereas the parallel formulation can be written as: y = x + MLP(LayerNorm(x)) + Attention(LayerNorm(x)) The parallel formulation results in roughly 15% faster training speed at large scales, since the MLP and Attention input matrix multiplications can be fused. Ablation experiments showed a small quality degradation at 8B scale but no quality degradation at 62B scale, so we extrapolated that the effect of parallel layers should be quality neutral at the 540B scale.

The Routing Transformer is a Transformer that endows self-attention with a sparse routing module based on online k-means. Each attention module considers a clustering of the space: the current timestep only attends to context belonging to the same cluster. In other word, the current time-step query is routed to a limited number of context through its cluster assignment.

SimVLM is a minimalist pretraining framework to reduce training complexity by exploiting large-scale weak supervision. It is trained end-to-end with a single prefix language modeling (PrefixLM) objective. PrefixLM enables bidirectional attention within the prefix sequence, and thus it is applicable for both decoder-only and encoder-decoder sequence-to-sequence language models.

Subformer is a Transformer that combines sandwich-style parameter sharing, which overcomes naive cross-layer parameter sharing in generative models, and self-attentive embedding factorization (SAFE). In SAFE, a small self-attention layer is used to reduce embedding parameter count.

TAPEX is a conceptually simple and empirically powerful pre-training approach to empower existing models with table reasoning skills. TAPEX realizes table pre-training by learning a neural SQL executor over a synthetic corpus, which is obtained by automatically synthesising executable SQL queries.

TSDAE is an unsupervised sentence embedding method. During training, TSDAE encodes corrupted sentences into fixed-sized vectors and requires the decoder to reconstruct the original sentences from this sentence embedding. For good reconstruction quality, the semantics must be captured well in the sentence embedding from the encoder. Later, at inference, we only use the encoder for creating sentence embeddings. The model architecture of TSDAE is a modified encoder-decoder Transformer where the key and value of the cross-attention are both confined to the sentence embedding only. Formally, the formulation of the modified cross-attention is: $$ H^{(k)}=\text { Attention }\left(H^{(k-1)},\left[s^{T}\right],\left[s^{T}\right]\right) $$ $$ \operatorname{Attention}(Q, K, V)=\operatorname{softmax}\left(\frac{Q K^{T}}{\sqrt{d}}\right) V $$ where $H^{(k)} \in \mathbb{R}^{t \times d}$ is the decoder hidden states within $t$ decoding steps at the $k$-th layer, $d$ is the size of the sentence embedding, $\left[s^{T}\right] \in \mathbb{R}^{1 \times d}$ is a one-row matrix including the sentence embedding vector and $Q, K$ and $V$ are the query, key and value, respectively. By exploring different configurations on the STS benchmark dataset, the authors discover that the best combination is: (1) adopting deletion as the input noise and setting the deletion ratio to $0.6,(2)$ using the output of the [CLS] token as fixed-sized sentence representation (3) tying the encoder and decoder parameters during training.

Vokenization is an approach for extrapolating multimodal alignments to language-only data by contextually mapping language tokens to their related images ("vokens") by retrieval. Instead of directly supervising the language model with visually grounded language datasets (e.g., MS COCO) these relative small datasets are used to train the vokenization processor (i.e. the vokenizer). Vokens are generated for large language corpora (e.g., English Wikipedia), and the visually-supervised language model takes the input supervision from these large datasets, thus bridging the gap between different data sources.

Please enter a description about the method here

AdaFisher – an adaptive second-order optimizer that leverages a block-diagonal approximation to the Fisher information matrix for adaptive gradient preconditioning.

An All-Attention Layer is an attention module and layer for transformers that merges the self-attention and feedforward sublayers into a single unified attention layer. As opposed to the two-step mechanism of the Transformer layer, it directly builds its representation from the context and a persistent memory block without going through a feedforward transformation. The additional persistent memory block stores, in the form of key-value vectors, information that does not depend on the context. In terms of parameters, these persistent key-value vectors replace the feedforward sublayer.

In this work, we present a conceptually simple and effective method to train a strong bilingual multimodal representation model. Starting from the pretrained multimodal representation model CLIP released by OpenAI, we switched its text encoder with a pretrained multilingual text encoder XLM-R, and aligned both languages and image representations by a two-stage training schema consisting of teacher learning and contrastive learning. We validate our method through evaluations of a wide range of tasks. We set new state-of-the-art performances on a bunch of tasks including ImageNet-CN, Flicker30k- CN, and COCO-CN. Further, we obtain very close performances with CLIP on almost all tasks, suggesting that one can simply alter the text encoder in CLIP for extended capabilities such as multilingual understanding. Our models and code are available at https://github.com/FlagAI-Open/FlagAI.

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.