Unlocking Continual Learning Abilities in Language Models

Language models (LMs) exhibit impressive performance and generalization capabilities. However, LMs struggle with the persistent challenge of catastrophic forgetting, which undermines their long-term sustainability in continual learning (CL). Existing approaches usually address the issue by incorporating old task data or task-wise inductive bias into LMs. However, old data and accurate task information are often unavailable or costly to collect, hindering the availability of current CL approaches for LMs. To address this limitation, we introduce $\textbf{MIGU}$ ($\textbf{M}$agn$\textbf{I}$tude-based $\textbf{G}$radient $\textbf{U}$pdating for continual learning), a rehearsal-free and task-label-free method that only updates the model parameters with large magnitudes of output in LMs' linear layers. MIGU is based on our observation that the L1-normalized magnitude distribution of the output in LMs' linear layers is different when the LM models deal with different task data. By imposing this simple constraint on the gradient update process, we can leverage the inherent behaviors of LMs, thereby unlocking their innate CL abilities. Our experiments demonstrate that MIGU is universally applicable to all three LM architectures (T5, RoBERTa, and Llama2), delivering state-of-the-art or on-par performance across continual finetuning and continual pre-training settings on four CL benchmarks. For example, MIGU brings a 15.2% average accuracy improvement over conventional parameter-efficient finetuning baselines in a 15-task CL benchmark. MIGU can also seamlessly integrate with all three existing CL types to further enhance performance. Code is available at \href{https://github.com/wenyudu/MIGU}{this https URL}.

Stacking Your Transformers: A Closer Look at Model Growth for Efficient LLM Pre-Training

LLMs are computationally expensive to pre-train due to their large scale. Model growth emerges as a promising approach by leveraging smaller models to accelerate the training of larger ones. However, the viability of these model growth methods in efficient LLM pre-training remains underexplored. This work identifies three critical $\underline{\textit{O}}$bstacles: ($\textit{O}$1) lack of comprehensive evaluation, ($\textit{O}$2) untested viability for scaling, and ($\textit{O}$3) lack of empirical guidelines. To tackle $\textit{O}$1, we summarize existing approaches into four atomic growth operators and systematically evaluate them in a standardized LLM pre-training setting. Our findings reveal that a depthwise stacking operator, called $G_{\text{stack}}$, exhibits remarkable acceleration in training, leading to decreased loss and improved overall performance on eight standard NLP benchmarks compared to strong baselines. Motivated by these promising results, we conduct extensive experiments to delve deeper into $G_{\text{stack}}$ to address $\textit{O}$2 and $\textit{O}$3. For $\textit{O}$2 (untested scalability), our study shows that $G_{\text{stack}}$ is scalable and consistently performs well, with experiments up to 7B LLMs after growth and pre-training LLMs with 750B tokens. For example, compared to a conventionally trained 7B model using 300B tokens, our $G_{\text{stack}}$ model converges to the same loss with 194B tokens, resulting in a 54.6\% speedup. We further address $\textit{O}$3 (lack of empirical guidelines) by formalizing guidelines to determine growth timing and growth factor for $G_{\text{stack}}$, making it practical in general LLM pre-training. We also provide in-depth discussions and comprehensive ablation studies of $G_{\text{stack}}$. Our code and pre-trained model are available at $\href{https://llm-stacking.github.io/}{https://llm-stacking.github.io/}$.

m2mKD: Module-to-Module Knowledge Distillation for Modular Transformers

Modular neural architectures are gaining increasing attention due to their powerful capability for generalization and sample-efficient adaptation to new domains. However, training modular models, particularly in the early stages, poses challenges due to the optimization difficulties arising from their intrinsic sparse connectivity. Leveraging the knowledge from monolithic models, using techniques such as knowledge distillation, is likely to facilitate the training of modular models and enable them to integrate knowledge from multiple models pretrained on diverse sources. Nevertheless, conventional knowledge distillation approaches are not tailored to modular models and can fail when directly applied due to the unique architectures and the enormous number of parameters involved. Motivated by these challenges, we propose a general module-to-module knowledge distillation (m2mKD) method for transferring knowledge between modules. Our approach involves teacher modules split from a pretrained monolithic model, and student modules of a modular model. m2mKD separately combines these modules with a shared meta model and encourages the student module to mimic the behaviour of the teacher module. We evaluate the effectiveness of m2mKD on two distinct modular neural architectures: Neural Attentive Circuits (NACs) and Vision Mixture-of-Experts (V-MoE). By applying m2mKD to NACs, we achieve significant improvements in IID accuracy on Tiny-ImageNet (up to 5.6%) and OOD robustness on Tiny-ImageNet-R (up to 4.2%). On average, we observe a 1% gain in both ImageNet and ImageNet-R. The V-MoE-Base model trained using m2mKD also achieves 3.5% higher accuracy than end-to-end training on ImageNet. The experimental results demonstrate that our method offers a promising solution for connecting modular networks with pretrained monolithic models. Code is available at https://github.com/kamanphoebe/m2mKD.

f-Divergence Minimization for Sequence-Level Knowledge Distillation

Knowledge distillation (KD) is the process of transferring knowledge from a large model to a small one. It has gained increasing attention in the natural language processing community, driven by the demands of compressing ever-growing language models. In this work, we propose an f-DISTILL framework, which formulates sequence-level knowledge distillation as minimizing a generalized f-divergence function. We propose four distilling variants under our framework and show that existing SeqKD and ENGINE approaches are approximations of our f-DISTILL methods. We further derive step-wise decomposition for our f-DISTILL, reducing intractable sequence-level divergence to word-level losses that can be computed in a tractable manner. Experiments across four datasets show that our methods outperform existing KD approaches, and that our symmetric distilling losses can better force the student to learn from the teacher distribution.

Graphix-T5: Mixing Pre-Trained Transformers with Graph-Aware Layers for Text-to-SQL Parsing

The task of text-to-SQL parsing, which aims at converting natural language questions into executable SQL queries, has garnered increasing attention in recent years, as it can assist end users in efficiently extracting vital information from databases without the need for technical background. One of the major challenges in text-to-SQL parsing is domain generalization, i.e., how to generalize well to unseen databases. Recently, the pre-trained text-to-text transformer model, namely T5, though not specialized for text-to-SQL parsing, has achieved state-of-the-art performance on standard benchmarks targeting domain generalization. In this work, we explore ways to further augment the pre-trained T5 model with specialized components for text-to-SQL parsing. Such components are expected to introduce structural inductive bias into text-to-SQL parsers thus improving model's capacity on (potentially multi-hop) reasoning, which is critical for generating structure-rich SQLs. To this end, we propose a new architecture GRAPHIX-T5, a mixed model with the standard pre-trained transformer model augmented by some specially-designed graph-aware layers. Extensive experiments and analysis demonstrate the effectiveness of GRAPHIX-T5 across four text-to-SQL benchmarks: SPIDER, SYN, REALISTIC and DK. GRAPHIX-T5 surpass all other T5-based parsers with a significant margin, achieving new state-of-the-art performance. Notably, GRAPHIX-T5-large reach performance superior to the original T5-large by 5.7% on exact match (EM) accuracy and 6.6% on execution accuracy (EX). This even outperforms the T5-3B by 1.2% on EM and 1.5% on EX.

Optimizing Stock Option Forecasting with the Assembly of Machine Learning Models and Improved Trading Strategies

This paper introduced key aspects of applying Machine Learning (ML) models, improved trading strategies, and the Quasi-Reversibility Method (QRM) to optimize stock option forecasting and trading results. It presented the findings of the follow-up project of the research "Application of Convolutional Neural Networks with Quasi-Reversibility Method Results for Option Forecasting". First, the project included an application of Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM) networks to provide a novel way of predicting stock option trends. Additionally, it examined the dependence of the ML models by evaluating the experimental method of combining multiple ML models to improve prediction results and decision-making. Lastly, two improved trading strategies and simulated investing results were presented. The Binomial Asset Pricing Model with discrete time stochastic process analysis and portfolio hedging was applied and suggested an optimized investment expectation. These results can be utilized in real-life trading strategies to optimize stock option investment results based on historical data.

Application of Convolutional Neural Networks with Quasi-Reversibility Method Results for Option Forecasting

This paper presents a novel way to apply mathematical finance and machine learning (ML) to forecast stock options prices. Following results from the paper Quasi-Reversibility Method and Neural Network Machine Learning to Solution of Black-Scholes Equations (appeared on the AMS Contemporary Mathematics journal), we create and evaluate new empirical mathematical models for the Black-Scholes equation to analyze data for 92,846 companies. We solve the Black-Scholes (BS) equation forwards in time as an ill-posed inverse problem, using the Quasi-Reversibility Method (QRM), to predict option price for the future one day. For each company, we have 13 elements including stock and option daily prices, volatility, minimizer, etc. Because the market is so complicated that there exists no perfect model, we apply ML to train algorithms to make the best prediction. The current stage of research combines QRM with Convolutional Neural Networks (CNN), which learn information across a large number of data points simultaneously. We implement CNN to generate new results by validating and testing on sample market data. We test different ways of applying CNN and compare our CNN models with previous models to see if achieving a higher profit rate is possible.

Linguistic dependencies and statistical dependence

What is the relationship between linguistic dependencies and statistical dependence? Building on earlier work in NLP and cognitive science, we study this question. We introduce a contextualized version of pointwise mutual information (CPMI), using pretrained language models to estimate probabilities of words in context. Extracting dependency trees which maximize CPMI, we compare the resulting structures against gold dependencies. Overall, we find that these maximum-CPMI trees correspond to linguistic dependencies more often than trees extracted from non-contextual PMI estimate, but only roughly as often as a simple baseline formed by connecting adjacent words. We also provide evidence that the extent to which the two kinds of dependency align cannot be explained by the distance between words or by the category of the dependency relation. Finally, our analysis sheds some light on the differences between large pretrained language models, specifically in the kinds of inductive biases they encode.

Exploiting Syntactic Structure for Better Language Modeling: A Syntactic Distance Approach

It is commonly believed that knowledge of syntactic structure should improve language modeling. However, effectively and computationally efficiently incorporating syntactic structure into neural language models has been a challenging topic. In this paper, we make use of a multi-task objective, i.e., the models simultaneously predict words as well as ground truth parse trees in a form called "syntactic distances", where information between these two separate objectives shares the same intermediate representation. Experimental results on the Penn Treebank and Chinese Treebank datasets show that when ground truth parse trees are provided as additional training signals, the model is able to achieve lower perplexity and induce trees with better quality.

GPSP: Graph Partition and Space Projection based Approach for Heterogeneous Network Embedding

In this paper, we propose GPSP, a novel Graph Partition and Space Projection based approach, to learn the representation of a heterogeneous network that consists of multiple types of nodes and links. Concretely, we first partition the heterogeneous network into homogeneous and bipartite subnetworks. Then, the projective relations hidden in bipartite subnetworks are extracted by learning the projective embedding vectors. Finally, we concatenate the projective vectors from bipartite subnetworks with the ones learned from homogeneous subnetworks to form the final representation of the heterogeneous network. Extensive experiments are conducted on a real-life dataset. The results demonstrate that GPSP outperforms the state-of-the-art baselines in two key network mining tasks: node classification and clustering.