Meta-Learning

GPT-4 is a transformer based model pre-trained to predict the next token in a document.

LLaMA is a collection of foundation language models ranging from 7B to 65B parameters. It is based on the transformer architecture with various improvements that were subsequently proposed. The main difference with the original architecture are listed below. - RMSNorm normalizing function is used to improve the training stability, by normalizing the input of each transformer sub-layer, instead of normalizing the output. - The ReLU non-linearity is replaced by the SwiGLU activation function to improve performance. - Absolute positional embeddings are removed and instead rotary positional embeddings (RoPE) are added at each layer of the network.

We present TabPFN, a trained Transformer that can do supervised classification for small tabular datasets in less than a second, needs no hyperparameter tuning and is competitive with state-of-the-art classification methods. TabPFN is fully entailed in the weights of our network, which accepts training and test samples as a set-valued input and yields predictions for the entire test set in a single forward pass. TabPFN is a Prior-Data Fitted Network (PFN) and is trained offline once, to approximate Bayesian inference on synthetic datasets drawn from our prior. This prior incorporates ideas from causal reasoning: It entails a large space of structural causal models with a preference for simple structures. On the 18 datasets in the OpenML-CC18 suite that contain up to 1 000 training data points, up to 100 purely numerical features without missing values, and up to 10 classes, we show that our method clearly outperforms boosted trees and performs on par with complex state-of-the-art AutoML systems with up to 230× speedup. This increases to a 5 700× speedup when using a GPU. We also validate these results on an additional 67 small numerical datasets from OpenML. We provide all our code, the trained TabPFN, an interactive browser demo and a Colab notebook at https://github.com/automl/TabPFN.

Meta Face Recognition (MFR) is a meta-learning face recognition method. MFR synthesizes the source/target domain shift with a meta-optimization objective, which requires the model to learn effective representations not only on synthesized source domains but also on synthesized target domains. Specifically, domain-shift batches are built through a domain-level sampling strategy and back-propagated gradients/meta-gradients are obtained on synthesized source/target domains by optimizing multi-domain distributions. The gradients and meta-gradients are further combined to update the model to improve generalization.

Latent Optimisation is a technique used for generative adversarial networks to refine the sample quality of $z$. Specifically, it exploits knowledge from the discriminator $D$ to refine the latent source $z$. Intuitively, the gradient $\nabla\_{z}f\left(z\right) = \delta{f}\left(z\right)\delta{z}$ points in the direction that better satisfies the discriminator $D$, which implies better samples. Therefore, instead of using the randomly sampled $z \sim p\left(z\right)$, we uses the optimised latent: $$ \Delta{z} = \alpha\frac{\delta{f}\left(z\right)}{\delta{z}} $$ $$ z' = z + \Delta{z} $$ Source: LOGAN .

The Simple Neural Attention Meta-Learner, or SNAIL, combines the benefits of temporal convolutions and attention to solve meta-learning tasks. They introduce positional dependence through temporal convolutions to make the model applicable to reinforcement tasks - where the observations, actions, and rewards are intrinsically sequential. They also introduce attention in order to provide pinpoint access over an infinitely large context. SNAIL is constructing by combining the two: we use temporal convolutions to produce the context over which we use a causal attention operation.

Euclidean Norm Regularization is a regularization step used in generative adversarial networks, and is typically added to both the generator and discriminator losses: $$ R\_{z} = w\_{r} \cdot ||\Delta{z}||^{2}\_{2} $$ where the scalar weight $w\_{r}$ is a parameter. Image: LOGAN

FIERCE is an entropic regularization on the feature space

Meta Pseudo Labels is a semi-supervised learning method that uses a teacher network to generate pseudo labels on unlabeled data to teach a student network. The teacher receives feedback from the student to inform the teacher to generate better pseudo labels. This feedback signal is used as a reward to train the teacher throughout the course of the student’s learning.

Meta-augmentation helps generate more varied tasks for a single example in meta-learning. It can be distinguished from data augmentation in classic machine learning as follows. For data augmentation in classical machine learning, the aim is to generate more varied examples, within a single task. Meta-augmentation has the exact opposite aim: we wish to generate more varied tasks, for a single example, to force the learner to quickly learn a new task from feedback. In meta-augmentation, adding randomness discourages the base learner and model from learning trivial solutions that do not generalize to new tasks.

CS-GAN is a type of generative adversarial network that uses a form of deep compressed sensing, and latent optimisation, to improve the quality of generated samples.

Kernel Inducing Points, or KIP, is a meta-learning algorithm for learning datasets that can mitigate the challenges which occur for naturally occurring datasets without a significant sacrifice in performance. KIP uses kernel-ridge regression to learn $\epsilon$-approximate datasets. It can be regarded as an adaption of the inducing point method for Gaussian processes to the case of Kernel Ridge Regression.

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

• 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.

AutoTinyBERT is a an efficient BERT variant found through neural architecture search. Specifically, one-shot learning is used to obtain a big Super Pretrained Language Model (SuperPLM), where the objectives of pre-training or task-agnostic BERT distillation are used. Then, given a specific latency constraint, an evolutionary algorithm is run on the SuperPLM to search optimal architectures. Finally, we extract the corresponding sub-models based on the optimal architectures and further train these models.

Base boosting is a generalization of gradient boosting, which fits a hybrid additive and varying coefficient model. - Namely, gradient boosting fits an additive model: \begin{equation} h(X ; { \alpha, \theta}) = \alpha_{0} + \sum_{k=1}^{K} \alpha_{k} b(X ; \theta_{k}), \end{equation} where the boosting mechanism begins optimization in function space at a constant model. - In contrast, base boosting fits the hybrid additive and varying coefficient model: \begin{equation} h(X ; { \alpha, \theta}) = \gamma(X) + \sum_{k=1}^{K} \alpha_{k} b(X ; \theta_{k}), \end{equation} where the boosting mechanism begins optimization in function space at a base model, which may be a non-constant model. A special case is the coordinate functional: \begin{equation} \gamma(X) = \pi_{j}(X) = X_{j} \end{equation} where $X_{j}$ denotes a prediction generated by the base model. - This setup facilitates knowledge transfer between the base model and boosting mechanism.

Fast Sample Re-Weighting, or FSR, is a sample re-weighting strategy to tackle problems such as dataset biases, noisy labels and imbalanced classes. It leverages a dictionary (essentially an extra buffer) to monitor the training history reflected by the model updates during meta optimization periodically, and utilises a valuation function to discover meaningful samples from training data as the proxy of reward data. The unbiased dictionary keeps being updated and provides reward signals to optimize sample weights. Additionally, instead of maintaining model states for both model and sample weight updates separately, feature sharing is enabled for saving the computation cost used for maintaining respective states.

MT-PET is a multi-task version of Pattern Exploiting Training (PET) for exaggeration detection, which leverages knowledge from complementary cloze-style QA tasks to improve few-shot learning. It defines pairs of complementary pattern-verbalizer pairs for a main task and auxiliary task. These PVPs are then used to train PET on data from both tasks. PET uses the masked language modeling objective of pretrained language models to transform a task into one or more cloze-style question answering tasks. In the original PET implementation, PVPs are defined for a single target task. MT-PET extends this by allowing for auxiliary PVPs from related tasks, adding complementary cloze-style QA tasks during training. The motivation for the multi-task approach is two-fold: 1) complementary cloze-style tasks can potentially help the model to learn different aspects of the main task, i.e. the similar tasks of exaggeration detection and claim strength prediction; 2) data on related tasks can be utilized during training, which is important in situations where data for the main task is limited.

SC-GPT is a multi-layer Transformer neural language model, trained in three steps: (i) Pre-trained on plain text, similar to GPT-2; (ii) Continuously pretrained on large amounts of dialog-act labeled utterances corpora to acquire the ability of controllable generation; (iii) Fine-tuned for a target domain using very limited amounts of domain labels. Unlike GPT-2, SC-GPT generates semantically controlled responses that are conditioned on the given semantic form, similar to SC-LSTM but requiring much less domain labels to generalize to new domains. It is pre-trained on a large set of annotated NLG corpus to acquire the controllable generation ability, and fine-tuned with only a few domain-specific labels to adapt to new domains.

STraTA, or Self-Training with Task Augmentation, is a self-training approach that builds on two key ideas for effective leverage of unlabeled data. First, STraTA uses task augmentation, a technique that synthesizes a large amount of data for auxiliary-task fine-tuning from target-task unlabeling texts. Second, STRATA performs self-training by further fine-tuning the strong base model created by task augmentation on a broad distribution of pseudo-labeled data. In task augmentation, we train an NLI data generation model and use it to synthesize a large amount of in-domain NLI training data for each given target task, which is then used for auxiliary (intermediate) fine-tuning. The self-training algorithm iteratively learns a better model using a concatenation of labeled and pseudo-labeled examples. At each iteration, we always start with the auxiliary-task model produced by task augmentation and train on a broad distribution of pseudo-labeled data.

SimAdapter is a module for explicitly learning knowledge from adapters. SimAdapter aims to learn the similarities between the source and target languages during fine-tuning using the adapters, and the similarity is based on an attention mechanism. The detailed composition of the SimAdapter is shown in the Figure. By taking the language-agnostic representations from the backbone model as the query, and the language-specific outputs from multiple adapter as the keys and values, the final output for SimAdapter over attention are computed as (For notation simplicity, we omit the layer index $l$ below): $$ \operatorname{SimAdapter}\left(\mathbf{z}, \mathbf{a}\_{\left\(S\_{1}, S\_{2}, \ldots, S\_{N}\right\)}\right)=\sum_{i=1}^{N} \operatorname{Attn}\left(\mathbf{z}, \mathbf{a}\_{S\_{i}}\right) \cdot\left(\mathbf{a}\_{S\_{i}} \mathbf{W}\_{V}\right) $$ where SimAdapter $(\cdot)$ and $\operatorname{Attn}(\cdot)$ denotes the SimAdapter and attention operations, respectively. Specifically, the attention operation is computed as: $$ \operatorname{Attn}(\mathbf{z}, \mathbf{a})=\operatorname{Softmax}\left(\frac{\left(\mathbf{z} \mathbf{W}\_{Q}\right)\left(\mathbf{a} \mathbf{W}\_{K}\right)^{\top}}{\tau}\right) $$ where $\tau$ is the temperature coefficient, $\mathbf{W}\_{Q}, \mathbf{W}\_{K}, \mathbf{W}\_{V}$ are attention matrices. Note that while $\mathbf{W}\_{Q}, \mathbf{W}\_{K}$ are initialized randomly, $\mathbf{W}\_{V}$ is initialized with a diagonal of ones and the rest of the matrix with small weights $(1 e-6)$ to retain the adapter representations. Furthermore, a regularization term is introduced to avoid drastic feature changes: $$ \mathcal{L}\_{\mathrm{reg}}=\sum\_{i, j}\left(\left(\mathbf{I}\_{V}\right)\_{i, j}-\left(\mathbf{W}\_{V}\right)_{i, j}\right)^{2} $$ where $\mathbf{I}\_{V}$ is the identity matrix with the same size as $\mathbf{W}\_{V}$