Language Modelling

Discriminative Fine-Tuning is a fine-tuning strategy that is used for ULMFiT type models. Instead of using the same learning rate for all layers of the model, discriminative fine-tuning allows us to tune each layer with different learning rates. For context, the regular stochastic gradient descent (SGD) update of a model’s parameters $\theta$ at time step $t$ looks like the following (Ruder, 2016): $$ \theta\_{t} = \theta\_{t-1} − \eta\cdot\nabla\_{\theta}J\left(\theta\right)$$ where $\eta$ is the learning rate and $\nabla\_{\theta}J\left(\theta\right)$ is the gradient with regard to the model’s objective function. For discriminative fine-tuning, we split the parameters $\theta$ into {$\theta\_{1}, \ldots, \theta\_{L}$} where $\theta\_{l}$ contains the parameters of the model at the $l$-th layer and $L$ is the number of layers of the model. Similarly, we obtain {$\eta\_{1}, \ldots, \eta\_{L}$} where $\theta\_{l}$ where $\eta\_{l}$ is the learning rate of the $l$-th layer. The SGD update with discriminative finetuning is then: $$ \theta\_{t}^{l} = \theta\_{t-1}^{l} - \eta^{l}\cdot\nabla\_{\theta^{l}}J\left(\theta\right) $$ The authors find that empirically it worked well to first choose the learning rate $\eta^{L}$ of the last layer by fine-tuning only the last layer and using $\eta^{l-1}=\eta^{l}/2.6$ as the learning rate for lower layers.

A Gated Linear Unit, or GLU computes: $$ \mathrm{GLU}(a, b) = a \otimes \sigma(b) $$ It is used in natural language processing architectures, for example the Gated CNN, because here $\sigma(b)$ is the gate that control what information from $a$ is passed up to the following layer. Intuitively, for a language modeling task, the gating mechanism allows selection of words or features that are important for predicting the next word. The GLU also has non-linear capabilities, but has a linear path for the gradient so diminishes the vanishing gradient problem.

Foundation models, now powering most of the exciting applications in deep learning, are almost universally based on the Transformer architecture and its core attention module. Many subquadratic-time architectures such as linear attention, gated convolution and recurrent models, and structured state space models (SSMs) have been developed to address Transformers’ computational inefficiency on long sequences, but they have not performed as well as attention on important modalities such as language. We identify that a key weakness of such models is their inability to perform content-based reasoning, and make several improvements. First, simply letting the SSM parameters be functions of the input addresses their weakness with discrete modalities, allowing the model to selectively propagate or forget information along the sequence length dimension depending on the current token. Second, even though this change prevents the use of efficient convolutions, we design a hardware-aware parallel algorithm in recurrent mode. We integrate these selective SSMs into a simplified end-to-end neural network architecture without attention or even MLP blocks (Mamba). Mamba enjoys fast inference (5× higher throughput than Transformers) and linear scaling in sequence length, and its performance improves on real data up to million-length sequences. As a general sequence model backbone, Mamba achieves state-of-the-art performance across several modalities such as language, audio, and genomics. On language modeling, our Mamba-3B model outperforms Transformers of the same size and matches Transformers twice its size, both in pre-training and downstream evaluation.

OPT is a suite of decoder-only pre-trained transformers ranging from 125M to 175B parameters. The model uses an AdamW optimizer and weight decay of 0.1. It follows a linear learning rate schedule, warming up from 0 to the maximum learning rate over the first 2000 steps in OPT-175B, or over 375M tokens in the smaller models, and decaying down to 10% of the maximum LR over 300B tokens. The batch sizes range from 0.5M to 4M depending on the model size and is kept constant throughout the course of training.

A HyperNetwork is a network that generates weights for a main network. The behavior of the main network is the same with any usual neural network: it learns to map some raw inputs to their desired targets; whereas the hypernetwork takes a set of inputs that contain information about the structure of the weights and generates the weight for that layer.

Embeddings from Language Models, or ELMo, is a type of deep contextualized word representation that models both (1) complex characteristics of word use (e.g., syntax and semantics), and (2) how these uses vary across linguistic contexts (i.e., to model polysemy). Word vectors are learned functions of the internal states of a deep bidirectional language model (biLM), which is pre-trained on a large text corpus. A biLM combines both a forward and backward LM. ELMo jointly maximizes the log likelihood of the forward and backward directions. To add ELMo to a supervised model, we freeze the weights of the biLM and then concatenate the ELMo vector $\textbf{ELMO}^{task}_k$ with $\textbf{x}_k$ and pass the ELMO enhanced representation $[\textbf{x}_k; \textbf{ELMO}^{task}_k]$ into the task RNN. Here $\textbf{x}_k$ is a context-independent token representation for each token position. Image Source: here

Differentiable Architecture Search (DART) is a method for efficient architecture search. The search space is made continuous so that the architecture can be optimized with respect to its validation set performance through gradient descent.

Variational Dropout is a regularization technique based on dropout, but uses a variational inference grounded approach. In Variational Dropout, we repeat the same dropout mask at each time step for both inputs, outputs, and recurrent layers (drop the same network units at each time step). This is in contrast to ordinary Dropout where different dropout masks are sampled at each time step for the inputs and outputs alone.

BLOOM is a decoder-only Transformer language model that was trained on the ROOTS corpus, a dataset comprising hundreds of sources in 46 natural and 13 programming languages (59 in total).

Performer is a Transformer architecture which can estimate regular (softmax) full-rank-attention Transformers with provable accuracy, but using only linear (as opposed to quadratic) space and time complexity, without relying on any priors such as sparsity or low-rankness. Performers are linear architectures fully compatible with regular Transformers and with strong theoretical guarantees: unbiased or nearly-unbiased estimation of the attention matrix, uniform convergence and low estimation variance. To approximate softmax attention-kernels, Performers use a Fast Attention Via positive Orthogonal Random features approach (FAVOR+), leveraging new methods for approximating softmax and Gaussian kernels.

A Highway Network is an architecture designed to ease gradient-based training of very deep networks. They allow unimpeded information flow across several layers on "information highways". The architecture is characterized by the use of gating units which learn to regulate the flow of information through a network. Highway networks with hundreds of layers can be trained directly using stochastic gradient descent and with a variety of activation functions.

Longformer is a modified Transformer architecture. Traditional Transformer-based models are unable to process long sequences due to their self-attention operation, which scales quadratically with the sequence length. To address this, Longformer uses an attention pattern that scales linearly with sequence length, making it easy to process documents of thousands of tokens or longer. The attention mechanism is a drop-in replacement for the standard self-attention and combines a local windowed attention with a task motivated global attention. The attention patterns utilised include: sliding window attention, dilated sliding window attention and global + sliding window. These can be viewed in the components section of this page.

Chain-of-thought prompts contain a series of intermediate reasoning steps, and they are shown to significantly improve the ability of large language models to perform certain tasks that involve complex reasoning (e.g., arithmetic, commonsense reasoning, symbolic reasoning, etc.)

DeepWalk learns embeddings (social representations) of a graph's vertices, by modeling a stream of short random walks. Social representations are latent features of the vertices that capture neighborhood similarity and community membership. These latent representations encode social relations in a continuous vector space with a relatively small number of dimensions. It generalizes neural language models to process a special language composed of a set of randomly-generated walks. The goal is to learn a latent representation, not only a probability distribution of node co-occurrences, and so as to introduce a mapping function $\Phi \colon v \in V \mapsto \mathbb{R}^{|V|\times d}$. This mapping $\Phi$ represents the latent social representation associated with each vertex $v$ in the graph. In practice, $\Phi$ is represented by a $|V| \times d$ matrix of free parameters.

Adaptive Input Embeddings extend the adaptive softmax to input word representations. The factorization assigns more capacity to frequent words and reduces the capacity for less frequent words with the benefit of reducing overfitting to rare words.

Embedding Dropout is equivalent to performing dropout on the embedding matrix at a word level, where the dropout is broadcast across all the word vector’s embedding. The remaining non-dropped-out word embeddings are scaled by $\frac{1}{1-p\_{e}}$ where $p\_{e}$ is the probability of embedding dropout. As the dropout occurs on the embedding matrix that is used for a full forward and backward pass, this means that all occurrences of a specific word will disappear within that pass, equivalent to performing variational dropout on the connection between the one-hot embedding and the embedding lookup. Source: Merity et al, Regularizing and Optimizing LSTM Language Models

Transformer-XL (meaning extra long) is a Transformer architecture that introduces the notion of recurrence to the deep self-attention network. Instead of computing the hidden states from scratch for each new segment, Transformer-XL reuses the hidden states obtained in previous segments. The reused hidden states serve as memory for the current segment, which builds up a recurrent connection between the segments. As a result, modeling very long-term dependency becomes possible because information can be propagated through the recurrent connections. As an additional contribution, the Transformer-XL uses a new relative positional encoding formulation that generalizes to attention lengths longer than the one observed during training.

Pythia is a suite of decoder-only autoregressive language models all trained on public data seen in the exact same order and ranging in size from 70M to 12B parameters. The model architecture and hyperparameters largely follow GPT-3, with a few notable deviations based on recent advances in best practices for large scale language modeling.

XLM is a Transformer based architecture that is pre-trained using one of three language modelling objectives: 1. Causal Language Modeling - models the probability of a word given the previous words in a sentence. 2. Masked Language Modeling - the masked language modeling objective of BERT. 3. Translation Language Modeling - a (new) translation language modeling objective for improving cross-lingual pre-training. The authors find that both the CLM and MLM approaches provide strong cross-lingual features that can be used for pretraining models.

Activation Regularization (AR), or $L\_{2}$ activation regularization, is regularization performed on activations as opposed to weights. It is usually used in conjunction with RNNs. It is defined as: $$\alpha{L}\_{2}\left(m\circ{h\_{t}}\right) $$ where $m$ is a dropout mask used by later parts of the model, $L\_{2}$ is the $L\_{2}$ norm, and $h_{t}$ is the output of an RNN at timestep $t$, and $\alpha$ is a scaling coefficient. When applied to the output of a dense layer, AR penalizes activations that are substantially away from 0, encouraging activations to remain small.

ERNIE is a transformer-based model consisting of two stacked modules: 1) textual encoder and 2) knowledgeable encoder, which is responsible to integrate extra token-oriented knowledge information into textual information. This layer consists of stacked aggregators, designed for encoding both tokens and entities as well as fusing their heterogeneous features. To integrate this layer of enhancing representations via knowledge, a special pre-training task is adopted for ERNIE - it involves randomly masking token-entity alignments and training the model to predict all corresponding entities based on aligned tokens (aka denoising entity auto-encoder).

ASGD Weight-Dropped LSTM, or AWD-LSTM, is a type of recurrent neural network that employs DropConnect for regularization, as well as NT-ASGD for optimization - non-monotonically triggered averaged SGD - which returns an average of last iterations of weights. Additional regularization techniques employed include variable length backpropagation sequences, variational dropout, embedding dropout, weight tying, independent embedding/hidden size, activation regularization and temporal activation regularization.

GLM is a bilingual (English and Chinese) pre-trained transformer-based language model that follow the traditional architecture of decoder-only autoregressive language modeling. It leverages autoregressive blank infilling as its training objective.

A new kind of implicit models, where the output of the network is defined as the solution to an "infinite-level" fixed point equation. Thanks to this we can compute the gradient of the output without activations and therefore with a significantly reduced memory footprint.

LXMERT is a model for learning vision-and-language cross-modality representations. It consists of a Transformer model that consists three encoders: object relationship encoder, a language encoder, and a cross-modality encoder. The model takes two inputs: image with its related sentence. The images are represented as a sequence of objects, whereas each sentence is represented as sequence of words. By combining the self-attention and cross-attention layers the model is able to generated language representation, image representations, and cross-modality representations from the input. The model is pre-trained with image-sentence pairs via five pre-training tasks: masked language modeling, masked object prediction, cross-modality matching, and image questions answering. These tasks help the model to learn both intra-modality and cross-modality relationships.

Slanted Triangular Learning Rates (STLR) is a learning rate schedule which first linearly increases the learning rate and then linearly decays it, which can be seen in Figure to the right. It is a modification of Triangular Learning Rates, with a short increase and a long decay period.

Universal Language Model Fine-tuning, or ULMFiT, is an architecture and transfer learning method that can be applied to NLP tasks. It involves a 3-layer AWD-LSTM architecture for its representations. The training consists of three steps: 1) general language model pre-training on a Wikipedia-based text, 2) fine-tuning the language model on a target task, and 3) fine-tuning the classifier on the target task. As different layers capture different types of information, they are fine-tuned to different extents using discriminative fine-tuning. Training is performed using Slanted triangular learning rates (STLR), a learning rate scheduling strategy that first linearly increases the learning rate and then linearly decays it. Fine-tuning the target classifier is achieved in ULMFiT using gradual unfreezing. Rather than fine-tuning all layers at once, which risks catastrophic forgetting, ULMFiT gradually unfreezes the model starting from the last layer (i.e., closest to the output) as this contains the least general knowledge. First the last layer is unfrozen and all unfrozen layers are fine-tuned for one epoch. Then the next group of frozen layers is unfrozen and fine-tuned and repeat, until all layers are fine-tuned until convergence at the last iteration.

In image inpainting task, the mechanism extracts complementary features from the word embedding in two paths by reciprocal attention, which is done by comparing the descriptive text and complementary image areas through reciprocal attention.

In this work, we pursue a unified paradigm for multimodal pretraining to break the scaffolds of complex task/modality-specific customization. We propose OFA, a Task-Agnostic and Modality-Agnostic framework that supports Task Comprehensiveness. OFA unifies a diverse set of cross-modal and unimodal tasks, including image generation, visual grounding, image captioning, image classification, language modeling, etc., in a simple sequence-to-sequence learning framework. OFA follows the instruction-based learning in both pretraining and finetuning stages, requiring no extra task-specific layers for downstream tasks. In comparison with the recent state-of-the-art vision & language models that rely on extremely large cross-modal datasets, OFA is pretrained on only 20M publicly available image-text pairs. Despite its simplicity and relatively small-scale training data, OFA achieves new SOTAs in a series of cross-modal tasks while attaining highly competitive performances on uni-modal tasks. Our further analysis indicates that OFA can also effectively transfer to unseen tasks and unseen domains. Our code and models are publicly available at https://github.com/OFA-Sys/OFA.

Zoneout is a method for regularizing RNNs. At each timestep, zoneout stochastically forces some hidden units to maintain their previous values. Like dropout, zoneout uses random noise to train a pseudo-ensemble, improving generalization. But by preserving instead of dropping hidden units, gradient information and state information are more readily propagated through time, as in feedforward stochastic depth networks.