Li Du

Xiang Li, Yiqun Yao, Xin Jiang et al. · tencent-ai, tsinghua

Large language models (LLMs) are considered important approaches towards foundational machine intelligence, achieving remarkable success in Natural Language Processing and multimodal tasks, among others. However, the carbon footprints and financial costs originating from heavy pre-training computation is a non-negligible issue. Progressive training methods, inspired by the neurogenesis process that grows neural structures, have shown potential to accelerate LLM pre-training. However, the algorithms, implementation, and practices for progressively training LLMs beyond 100B parameters remain underexplored. In this paper, we show that our model, namely FLM-101B, trained with our growth strategy under a budget of \$100K, reaches 80\% of the baselines' performances with only 10\% of their floating-point operations. We believe that further studies on progressive training will benefit the community by cutting down the costs and promoting green AI. The checkpoint of FLM-101B is released at https://huggingface.co/CofeAI/FLM-101B.

Yijiang Liu, Huanrui Yang, Zhen Dong et al. · berkeley

The complicated architecture and high training cost of vision transformers urge the exploration of post-training quantization. However, the heavy-tailed distribution of vision transformer activations hinders the effectiveness of previous post-training quantization methods, even with advanced quantizer designs. Instead of tuning the quantizer to better fit the complicated activation distribution, this paper proposes NoisyQuant, a quantizer-agnostic enhancement for the post-training activation quantization performance of vision transformers. We make a surprising theoretical discovery that for a given quantizer, adding a fixed Uniform noisy bias to the values being quantized can significantly reduce the quantization error under provable conditions. Building on the theoretical insight, NoisyQuant achieves the first success on actively altering the heavy-tailed activation distribution with additive noisy bias to fit a given quantizer. Extensive experiments show NoisyQuant largely improves the post-training quantization performance of vision transformer with minimal computation overhead. For instance, on linear uniform 6-bit activation quantization, NoisyQuant improves SOTA top-1 accuracy on ImageNet by up to 1.7%, 1.1% and 0.5% for ViT, DeiT, and Swin Transformer respectively, achieving on-par or even higher performance than previous nonlinear, mixed-precision quantization.

Lirui Xiao, Huanrui Yang, Zhen Dong et al. · berkeley

Mixed-precision quantization has been widely applied on deep neural networks (DNNs) as it leads to significantly better efficiency-accuracy tradeoffs compared to uniform quantization. Meanwhile, determining the exact precision of each layer remains challenging. Previous attempts on bit-level regularization and pruning-based dynamic precision adjustment during training suffer from noisy gradients and unstable convergence. In this work, we propose Continuous Sparsification Quantization (CSQ), a bit-level training method to search for mixed-precision quantization schemes with improved stability. CSQ stabilizes the bit-level mixed-precision training process with a bi-level gradual continuous sparsification on both the bit values of the quantized weights and the bit selection in determining the quantization precision of each layer. The continuous sparsification scheme enables fully-differentiable training without gradient approximation while achieving an exact quantized model in the end.A budget-aware regularization of total model size enables the dynamic growth and pruning of each layer's precision towards a mixed-precision quantization scheme of the desired size. Extensive experiments show CSQ achieves better efficiency-accuracy tradeoff than previous methods on multiple models and datasets.

Li Du, Lucas Torroba Hennigen, Tiago Pimentel et al. · cambridge, microsoft-research

Language modeling, a central task in natural language processing, involves estimating a probability distribution over strings. In most cases, the estimated distribution sums to 1 over all finite strings. However, in some pathological cases, probability mass can ``leak'' onto the set of infinite sequences. In order to characterize the notion of leakage more precisely, this paper offers a measure-theoretic treatment of language modeling. We prove that many popular language model families are in fact tight, meaning that they will not leak in this sense. We also generalize characterizations of tightness proposed in previous works.

Qingyuan Li, Yifan Zhang, Liang Li et al.

In the era of large-scale language models, the substantial parameter size poses significant challenges for deployment. Being a prevalent compression technique, quantization has emerged as the mainstream practice to tackle this issue, which is mainly centered on two recipes W8A8 and W4A16 (i.e. weights and activations in such bit widths). In this study, we propose a novel W4A8 post-training quantization method for the available open-sourced LLMs, which combines the advantages of both two recipes. Therefore, we can leverage the benefit in the I/O utilization of 4-bit weight quantization and the acceleration due to 8-bit matrix computation. Nevertheless, the W4A8 faces notorious performance degradation. As a remedy, we involve layerwise activation quantization strategies which feature a novel logarithmic equalization for most intractable layers, and we combine them with fine-grained weight quantization. Without whistles and bells, we eliminate the necessity for further fine-tuning and obtain the state-of-the-art W4A8 quantized performance on BLOOM, LLaMA, and LLaMA-2 on standard benchmarks. We confirm that the W4A8 quantization is achievable for the deployment of large language models, fostering their wide-spreading real-world applications.

Vilém Zouhar, Clara Meister, Juan Luis Gastaldi et al. · eth-zurich

Subword tokenization is a key part of many NLP pipelines. However, little is known about why some tokenizer and hyperparameter combinations lead to better downstream model performance than others. We propose that good tokenizers lead to \emph{efficient} channel usage, where the channel is the means by which some input is conveyed to the model and efficiency can be quantified in information-theoretic terms as the ratio of the Shannon entropy to the maximum possible entropy of the token distribution. Yet, an optimal encoding according to Shannon entropy assigns extremely long codes to low-frequency tokens and very short codes to high-frequency tokens. Defining efficiency in terms of Rényi entropy, on the other hand, penalizes distributions with either very high or very low-frequency tokens. In machine translation, we find that across multiple tokenizers, the Rényi entropy with $α= 2.5$ has a very strong correlation with \textsc{Bleu}: $0.78$ in comparison to just $-0.32$ for compressed length.

Li Du, Xiao Ding, Kai Xiong et al.

Understanding causality has vital importance for various Natural Language Processing (NLP) applications. Beyond the labeled instances, conceptual explanations of the causality can provide deep understanding of the causal facts to facilitate the causal reasoning process. However, such explanation information still remains absent in existing causal reasoning resources. In this paper, we fill this gap by presenting a human-annotated explainable CAusal REasoning dataset (e-CARE), which contains over 21K causal reasoning questions, together with natural language formed explanations of the causal questions. Experimental results show that generating valid explanations for causal facts still remains especially challenging for the state-of-the-art models, and the explanation information can be helpful for promoting the accuracy and stability of causal reasoning models.

Jianing Li, Jiaming Liu, Xiaobao Wei et al. · pku

Depth estimation is essential for various important real-world applications such as autonomous driving. However, it suffers from severe performance degradation in high-velocity scenario since traditional cameras can only capture blurred images. To deal with this problem, the spike camera is designed to capture the pixel-wise luminance intensity at high frame rate. However, depth estimation with spike camera remains very challenging using traditional monocular or stereo depth estimation algorithms, which are based on the photometric consistency. In this paper, we propose a novel Uncertainty-Guided Depth Fusion (UGDF) framework to fuse the predictions of monocular and stereo depth estimation networks for spike camera. Our framework is motivated by the fact that stereo spike depth estimation achieves better results at close range while monocular spike depth estimation obtains better results at long range. Therefore, we introduce a dual-task depth estimation architecture with a joint training strategy and estimate the distributed uncertainty to fuse the monocular and stereo results. In order to demonstrate the advantage of spike depth estimation over traditional camera depth estimation, we contribute a spike-depth dataset named CitySpike20K, which contains 20K paired samples, for spike depth estimation. UGDF achieves state-of-the-art results on CitySpike20K, surpassing all monocular or stereo spike depth estimation baselines. We conduct extensive experiments to evaluate the effectiveness and generalization of our method on CitySpike20K. To the best of our knowledge, our framework is the first dual-task fusion framework for spike camera depth estimation. Code and dataset will be released.

Ryan Cotterell, Anej Svete, Clara Meister et al. · allen-ai, eth-zurich

Large language models have become one of the most commonly deployed NLP inventions. In the past half-decade, their integration into core natural language processing tools has dramatically increased the performance of such tools, and they have entered the public discourse surrounding artificial intelligence. Consequently, it is important for both developers and researchers alike to understand the mathematical foundations of large language models, as well as how to implement them. These notes are the accompaniment to the theoretical portion of the ETH Zürich course on large language models, covering what constitutes a language model from a formal, theoretical perspective.

Li Du, Xiao Ding, Yue Zhang et al.

Predicting the subsequent event for an existing event context is an important but challenging task, as it requires understanding the underlying relationship between events. Previous methods propose to retrieve relational features from event graph to enhance the modeling of event correlation. However, the sparsity of event graph may restrict the acquisition of relevant graph information, and hence influence the model performance. To address this issue, we consider automatically building of event graph using a BERT model. To this end, we incorporate an additional structured variable into BERT to learn to predict the event connections in the training process. Hence, in the test process, the connection relationship for unseen events can be predicted by the structured variable. Results on two event prediction tasks: script event prediction and story ending prediction, show that our approach can outperform state-of-the-art baseline methods.

Tingting Wu, Xiao Ding, Hao Zhang et al.

Given data with label noise (i.e., incorrect data), deep neural networks would gradually memorize the label noise and impair model performance. To relieve this issue, curriculum learning is proposed to improve model performance and generalization by ordering training samples in a meaningful (e.g., easy to hard) sequence. Previous work takes incorrect samples as generic hard ones without discriminating between hard samples (i.e., hard samples in correct data) and incorrect samples. Indeed, a model should learn from hard samples to promote generalization rather than overfit to incorrect ones. In this paper, we address this problem by appending a novel loss function DiscrimLoss, on top of the existing task loss. Its main effect is to automatically and stably estimate the importance of easy samples and difficult samples (including hard and incorrect samples) at the early stages of training to improve the model performance. Then, during the following stages, DiscrimLoss is dedicated to discriminating between hard and incorrect samples to improve the model generalization. Such a training strategy can be formulated dynamically in a self-supervised manner, effectively mimicking the main principle of curriculum learning. Experiments on image classification, image regression, text sequence regression, and event relation reasoning demonstrate the versatility and effectiveness of our method, particularly in the presence of diversified noise levels.

Kai Xiong, Xiao Ding, Zhongyang Li et al.

Causal chain reasoning (CCR) is an essential ability for many decision-making AI systems, which requires the model to build reliable causal chains by connecting causal pairs. However, CCR suffers from two main transitive problems: threshold effect and scene drift. In other words, the causal pairs to be spliced may have a conflicting threshold boundary or scenario. To address these issues, we propose a novel Reliable Causal chain reasoning framework~(ReCo), which introduces exogenous variables to represent the threshold and scene factors of each causal pair within the causal chain, and estimates the threshold and scene contradictions across exogenous variables via structural causal recurrent neural networks~(SRNN). Experiments show that ReCo outperforms a series of strong baselines on both Chinese and English CCR datasets. Moreover, by injecting reliable causal chain knowledge distilled by ReCo, BERT can achieve better performances on four downstream causal-related tasks than BERT models enhanced by other kinds of knowledge.

Shuo Xie, Jiahao Qiu, Ankita Pasad et al.

While transferring a pretrained language model, common approaches conventionally attach their task-specific classifiers to the top layer and adapt all the pretrained layers. We investigate whether one could make a task-specific selection on which subset of the layers to adapt and where to place the classifier. The goal is to reduce the computation cost of transfer learning methods (e.g. fine-tuning or adapter-tuning) without sacrificing its performance. We propose to select layers based on the variability of their hidden states given a task-specific corpus. We say a layer is already "well-specialized" in a task if the within-class variability of its hidden states is low relative to the between-class variability. Our variability metric is cheap to compute and doesn't need any training or hyperparameter tuning. It is robust to data imbalance and data scarcity. Extensive experiments on the GLUE benchmark demonstrate that selecting layers based on our metric can yield significantly stronger performance than using the same number of top layers and often match the performance of fine-tuning or adapter-tuning the entire language model.

Franz Nowak, Anej Svete, Li Du et al. · allen-ai, eth-zurich

This work investigates the computational expressivity of language models (LMs) based on recurrent neural networks (RNNs). Siegelmann and Sontag (1992) famously showed that RNNs with rational weights and hidden states and unbounded computation time are Turing complete. However, LMs define weightings over strings in addition to just (unweighted) language membership and the analysis of the computational power of RNN LMs (RLMs) should reflect this. We extend the Turing completeness result to the probabilistic case, showing how a rationally weighted RLM with unbounded computation time can simulate any deterministic probabilistic Turing machine (PTM) with rationally weighted transitions. Since, in practice, RLMs work in real-time, processing a symbol at every time step, we treat the above result as an upper bound on the expressivity of RLMs. We also provide a lower bound by showing that under the restriction to real-time computation, such models can simulate deterministic real-time rational PTMs.

Yang Zhao, Li Du, Xiao Ding et al.

LLMs' performance on complex tasks is still unsatisfactory. A key issue is that presently LLMs learn in a data-driven schema, while the instructions about these complex tasks are both scarce and hard to collect or construct. On the contrary, a prominent phenomenon is that LLMs can learn rather fast on simpler tasks with adequate prior knowledge captured during pretraining stage. Thus, if the prerequisite and mechanism of such rapid generalization could be elucidated, it could enhance the efficiency and effectiveness of the LLM's ability to learn complex tasks. Thus, in this paper, we employ a gradient-based method, to dissect the process that the SFT process adapts LLMs to downstream tasks via the perspective of attention patterns. We find that: (1) LLMs selectively activate task-specific attention heads during SFT; (2) activation patterns for complex tasks are combinations of basic task patterns; and (3) changes in a few parameters can significantly impact activation patterns after SFT on a small number of samples.Based on these insights, experiments are conducted to actually enhance the efficiency and effectiveness of SFT.

Xingrun Xing, Boyan Gao, Zheng Zhang et al.

Recent advancements in large language models (LLMs) with billions of parameters have improved performance in various applications, but their inference processes demand significant energy and computational resources. In contrast, the human brain, with approximately 86 billion neurons, is much more energy-efficient than LLMs with similar parameters. Inspired by this, we redesign 7$\sim$70 billion parameter LLMs using bio-plausible spiking mechanisms, emulating the efficient behavior of the human brain. We propose the first spiking large language model, SpikeLLM. Coupled with the proposed model, two essential approaches are proposed to improve spike training efficiency: Generalized Integrate-and-Fire (GIF) neurons to compress spike length from $T$ to $\frac{T}{L} \log_2 L$ bits, and an Optimal Brain Spiking framework to divide outlier channels and allocate different $T$ for GIF neurons, which further compresses spike length to approximate $log_2T$ bits. The necessity of spike-driven LLM is proved by comparison with quantized LLMs with similar operations. In the OmniQuant pipeline, SpikeLLM reduces 11.01% WikiText2 perplexity and improves 2.55% accuracy of common scene reasoning on a LLAMA-7B W4A4 model. In the GPTQ pipeline, SpikeLLM achieves direct additive in linear layers, significantly exceeding PB-LLMs.

Afra Amini, Li Du, Ryan Cotterell

Gradient-based sampling algorithms have demonstrated their effectiveness in text generation, especially in the context of controlled text generation. However, there exists a lack of theoretically grounded and principled approaches for this task. In this paper, we take an important step toward building a principled approach for sampling from language models with gradient-based methods. We use discrete distributions given by language models to define densities and develop an algorithm based on Hamiltonian Monte Carlo to sample from them. We name our gradient-based technique Structured Voronoi Sampling (SVS). In an experimental setup where the reference distribution is known, we show that the empirical distribution of SVS samples is closer to the reference distribution compared to alternative sampling schemes. Furthermore, in a controlled generation task, SVS is able to generate fluent and diverse samples while following the control targets significantly better than other methods.

Kai Xiong, Xiao Ding, Li Du et al.

Large Language Models (LLMs) are versatile and demonstrate impressive generalization ability by mining and learning information from extensive unlabeled text. However, they still exhibit reasoning mistakes, often stemming from knowledge deficiencies, which can affect their trustworthiness and reliability. Although users can provide diverse and comprehensive queries, obtaining sufficient and effective feedback is demanding. Furthermore, evaluating LLMs comprehensively with limited labeled samples is difficult. This makes it a challenge to diagnose and remedy the deficiencies of LLMs through rich label-free user queries. To tackle this challenge, we propose a label-free curricular meaningful learning framework (LaMer). LaMer first employs relative entropy to automatically diagnose and quantify the knowledge deficiencies of LLMs in a label-free setting. Next, to remedy the diagnosed knowledge deficiencies, we apply curricular meaningful learning: first, we adopt meaningful learning to adaptively synthesize augmentation data according to the severity of the deficiencies, and then design a curricular deficiency remedy strategy to remedy the knowledge deficiencies of LLMs progressively. Experiments show that LaMer efficiently and effectively diagnoses and remedies knowledge deficiencies in LLMs, improving various LLMs across seven out-of-distribution (OOD) reasoning and language understanding benchmarks, achieving comparable results to baselines with just 40\% training data. LaMer even surpasses methods that rely on labeled datasets for deficiency diagnosis. In application, our label-free method can offer an effective knowledge deficiency diagnostic tool for efficient LLM development.

Rongyu Zhang, Yijiang Liu, Huanrui Yang et al.

Large language models (LLMs) encounter significant adaptation challenges in diverse multitask finetuning. Mixture-of-experts (MoE) provides a promising solution with a dynamic architecture, enabling effective task decoupling. However, scaling up the number of MoE experts incurs substantial parameter and computational overheads and suffers from limited performance gain due to naive routing mechanisms. In this paper, we design a novel framework, mix\underline{\textbf{T}}ure\underline{\textbf{-}}of-\underline{\textbf{R}}ank-on\underline{\textbf{E}}-e\underline{\textbf{X}}perts (\texttt{T-REX}), which leverages the combination of ultra-low rank experts to construct LoRA weights on pretrained LLMs. The rank-1 experts enable a mix-and-match mechanism to quadratically expand the vector subspace of experts with linear parameter overheads, achieving approximate error reduction with optimal efficiency. In addition, T-REX offers implicit guidance to the router, leveraging the inherent semantic clustering of training embeddings as prior knowledge, enabling optimized feature allocation across experts for a smoother convergence. Extensive theoretical and empirical results demonstrate that T-REX achieves superior efficiency and generalizability across diverse tasks. Compared with other LoRA-based methods, T-REX achieves up to 1.78\% mean accuracy improvement with around 30\%-40\% less trainable parameters across 14 public datasets. \href{https://github.com/RoyZry98/T-REX-Pytorch}{Code} is available.

Benjamin Lipkin, Benjamin LeBrun, Jacob Hoover Vigly et al.

The dominant approach to generating from language models subject to some constraint is locally constrained decoding (LCD), incrementally sampling tokens at each time step such that the constraint is never violated. Typically, this is achieved through token masking: looping over the vocabulary and excluding non-conforming tokens. There are two important problems with this approach. (i) Evaluating the constraint on every token can be prohibitively expensive -- LM vocabularies often exceed $100,000$ tokens. (ii) LCD can distort the global distribution over strings, sampling tokens based only on local information, even if they lead down dead-end paths. This work introduces a new algorithm that addresses both these problems. First, to avoid evaluating a constraint on the full vocabulary at each step of generation, we propose an adaptive rejection sampling algorithm that typically requires orders of magnitude fewer constraint evaluations. Second, we show how this algorithm can be extended to produce low-variance, unbiased estimates of importance weights at a very small additional cost -- estimates that can be soundly used within previously proposed sequential Monte Carlo algorithms to correct for the myopic behavior of local constraint enforcement. Through extensive empirical evaluation in text-to-SQL, molecular synthesis, goal inference, pattern matching, and JSON domains, we show that our approach is superior to state-of-the-art baselines, supporting a broader class of constraints and improving both runtime and performance. Additional theoretical and empirical analyses show that our method's runtime efficiency is driven by its dynamic use of computation, scaling with the divergence between the unconstrained and constrained LM, and as a consequence, runtime improvements are greater for better models.

Li Du, Afra Amini, Lucas Torroba Hennigen et al. · mit

Recent papers have demonstrated the possibility of energy-based text generation by adapting gradient-based sampling algorithms, a paradigm of MCMC algorithms that promises fast convergence. However, as we show in this paper, previous attempts on this approach to text generation all fail to sample correctly from the target language model distributions. To address this limitation, we consider the problem of designing text samplers that are faithful, meaning that they have the target text distribution as its limiting distribution. We propose several faithful gradient-based sampling algorithms to sample from the target energy-based text distribution correctly, and study their theoretical properties. Through experiments on various forms of text generation, we demonstrate that faithful samplers are able to generate more fluent text while adhering to the control objectives better.

Li Du, Hongyuan Mei, Jason Eisner

To predict the next token, autoregressive models ordinarily examine the past. Could they also benefit from also examining hypothetical futures? We consider a novel Transformer-based autoregressive architecture that estimates the next-token distribution by extrapolating multiple continuations of the past, according to some proposal distribution, and attending to these extended strings. This architecture draws insights from classical AI systems such as board game players: when making a local decision, a policy may benefit from exploring possible future trajectories and analyzing them. On multiple tasks including morphological inflection and Boolean satisfiability, our lookahead model is able to outperform the ordinary Transformer model of comparable size. However, on some tasks, it appears to be benefiting from the extra computation without actually using the lookahead information. We discuss possible variant architectures as well as future speedups.

Guhong Nie, Lirui Xiao, Menglong Zhu et al.

For binary neural networks (BNNs) to become the mainstream on-device computer vision algorithm, they must achieve a superior speed-vs-accuracy tradeoff than 8-bit quantization and establish a similar degree of general applicability in vision tasks. To this end, we propose a BNN framework comprising 1) a minimalistic inference scheme for hardware-friendliness, 2) an over-parameterized training scheme for high accuracy, and 3) a simple procedure to adapt to different vision tasks. The resultant framework overtakes 8-bit quantization in the speed-vs-accuracy tradeoff for classification, detection, segmentation, super-resolution and matching: our BNNs not only retain the accuracy levels of their 8-bit baselines but also showcase 1.3-2.4$\times$ faster FPS on mobile CPUs. Similar conclusions can be drawn for prototypical systolic-array-based AI accelerators, where our BNNs promise 2.8-7$\times$ fewer execution cycles than 8-bit and 2.1-2.7$\times$ fewer cycles than alternative BNN designs. These results suggest that the time for large-scale BNN adoption could be upon us.