Tao Lei

Joshua Ainslie, Tao Lei, Michiel de Jong et al. · deepmind

Many natural language processing tasks benefit from long inputs, but processing long documents with Transformers is expensive -- not only due to quadratic attention complexity but also from applying feedforward and projection layers to every token. However, not all tokens are equally important, especially for longer documents. We propose CoLT5, a long-input Transformer model that builds on this intuition by employing conditional computation, devoting more resources to important tokens in both feedforward and attention layers. We show that CoLT5 achieves stronger performance than LongT5 with much faster training and inference, achieving SOTA on the long-input SCROLLS benchmark. Moreover, CoLT5 can effectively and tractably make use of extremely long inputs, showing strong gains up to 64k input length.

Tao Lei, Junwen Bai, Siddhartha Brahma et al. · deepmind

We propose Conditional Adapter (CoDA), a parameter-efficient transfer learning method that also improves inference efficiency. CoDA generalizes beyond standard adapter approaches to enable a new way of balancing speed and accuracy using conditional computation. Starting with an existing dense pretrained model, CoDA adds sparse activation together with a small number of new parameters and a light-weight training phase. Our experiments demonstrate that the CoDA approach provides an unexpectedly efficient way to transfer knowledge. Across a variety of language, vision, and speech tasks, CoDA achieves a 2x to 8x inference speed-up compared to the state-of-the-art Adapter approaches with moderate to no accuracy loss and the same parameter efficiency.

Tao Lei, Rui Sun, Xuan Wang et al.

The hybrid architecture of convolutional neural networks (CNNs) and Transformer are very popular for medical image segmentation. However, it suffers from two challenges. First, although a CNNs branch can capture the local image features using vanilla convolution, it cannot achieve adaptive feature learning. Second, although a Transformer branch can capture the global features, it ignores the channel and cross-dimensional self-attention, resulting in a low segmentation accuracy on complex-content images. To address these challenges, we propose a novel hybrid architecture of convolutional neural networks hand in hand with vision Transformers (CiT-Net) for medical image segmentation. Our network has two advantages. First, we design a dynamic deformable convolution and apply it to the CNNs branch, which overcomes the weak feature extraction ability due to fixed-size convolution kernels and the stiff design of sharing kernel parameters among different inputs. Second, we design a shifted-window adaptive complementary attention module and a compact convolutional projection. We apply them to the Transformer branch to learn the cross-dimensional long-term dependency for medical images. Experimental results show that our CiT-Net provides better medical image segmentation results than popular SOTA methods. Besides, our CiT-Net requires lower parameters and less computational costs and does not rely on pre-training. The code is publicly available at https://github.com/SR0920/CiT-Net.

Zexuan Zhong, Tao Lei, Danqi Chen · princeton

Recent work has improved language models (LMs) remarkably by equipping them with a non-parametric memory component. However, most existing approaches only introduce mem-ories at testing time or represent them using a separately trained encoder, resulting in suboptimal training of the language model. In this work, we present TRIME, a novel yet simple training approach designed for training LMs with memory augmentation. Our approach uses a training objective that directly takes in-batch examples as accessible memory. We also present new methods for memory construction and data batching, which are used for adapting to different sets of memories--local, long-term, and external memory--at testing time. We evaluate TRIME on multiple language modeling and machine translation benchmarks and show that it is able to achieve significant improvements across all the settings. Concretely, TRIME reduces the perplexity from 18.70 to 15.37 on WIKITEXT-103, by effectively leveraging a large memory set from the training corpus. Compared to standard LM training, TRIME adds negligible computational overhead and is compatible with different neural architectures, making it a versatile solution for training memory-augmented LMs.

Jinhyuk Lee, Zhuyun Dai, Sai Meher Karthik Duddu et al. · cmu

Multi-vector retrieval models such as ColBERT [Khattab and Zaharia, 2020] allow token-level interactions between queries and documents, and hence achieve state of the art on many information retrieval benchmarks. However, their non-linear scoring function cannot be scaled to millions of documents, necessitating a three-stage process for inference: retrieving initial candidates via token retrieval, accessing all token vectors, and scoring the initial candidate documents. The non-linear scoring function is applied over all token vectors of each candidate document, making the inference process complicated and slow. In this paper, we aim to simplify the multi-vector retrieval by rethinking the role of token retrieval. We present XTR, ConteXtualized Token Retriever, which introduces a simple, yet novel, objective function that encourages the model to retrieve the most important document tokens first. The improvement to token retrieval allows XTR to rank candidates only using the retrieved tokens rather than all tokens in the document, and enables a newly designed scoring stage that is two-to-three orders of magnitude cheaper than that of ColBERT. On the popular BEIR benchmark, XTR advances the state-of-the-art by 2.8 nDCG@10 without any distillation. Detailed analysis confirms our decision to revisit the token retrieval stage, as XTR demonstrates much better recall of the token retrieval stage compared to ColBERT.

Rui Sun, Tao Lei, Weichuan Zhang et al.

The hybrid architecture of convolution neural networks (CNN) and Transformer has been the most popular method for medical image segmentation. However, the existing networks based on the hybrid architecture suffer from two problems. First, although the CNN branch can capture image local features by using convolution operation, the vanilla convolution is unable to achieve adaptive extraction of image features. Second, although the Transformer branch can model the global information of images, the conventional self-attention only focuses on the spatial self-attention of images and ignores the channel and cross-dimensional self-attention leading to low segmentation accuracy for medical images with complex backgrounds. To solve these problems, we propose vision Transformer embrace convolutional neural networks for medical image segmentation (TEC-Net). Our network has two advantages. First, dynamic deformable convolution (DDConv) is designed in the CNN branch, which not only overcomes the difficulty of adaptive feature extraction using fixed-size convolution kernels, but also solves the defect that different inputs share the same convolution kernel parameters, effectively improving the feature expression ability of CNN branch. Second, in the Transformer branch, a (shifted)-window adaptive complementary attention module ((S)W-ACAM) and compact convolutional projection are designed to enable the network to fully learn the cross-dimensional long-range dependency of medical images with few parameters and calculations. Experimental results show that the proposed TEC-Net provides better medical image segmentation results than SOTA methods including CNN and Transformer networks. In addition, our TEC-Net requires fewer parameters and computational costs and does not rely on pre-training. The code is publicly available at https://github.com/SR0920/TEC-Net.

Shu Liu, Enquan Huang, Ziyu Zhou et al.

Facial attractiveness prediction (FAP) aims to assess facial attractiveness automatically based on human aesthetic perception. Previous methods using deep convolutional neural networks have improved the performance, but their large-scale models have led to a deficiency in flexibility. In addition, most methods fail to take full advantage of the dataset. In this paper, we present a novel end-to-end FAP approach that integrates dual label distribution and lightweight design. The manual ratings, attractiveness score, and standard deviation are aggregated explicitly to construct a dual-label distribution to make the best use of the dataset, including the attractiveness distribution and the rating distribution. Such distributions, as well as the attractiveness score, are optimized under a joint learning framework based on the label distribution learning (LDL) paradigm. The data processing is simplified to a minimum for a lightweight design, and MobileNetV2 is selected as our backbone. Extensive experiments are conducted on two benchmark datasets, where our approach achieves promising results and succeeds in balancing performance and efficiency. Ablation studies demonstrate that our delicately designed learning modules are indispensable and correlated. Additionally, the visualization indicates that our approach can perceive facial attractiveness and capture attractive facial regions to facilitate semantic predictions. The code is available at https://github.com/enquan/2D_FAP.

Yujie Qian, Jinhyuk Lee, Sai Meher Karthik Duddu et al. · cmu, deepmind

Multi-vector retrieval models improve over single-vector dual encoders on many information retrieval tasks. In this paper, we cast the multi-vector retrieval problem as sparse alignment between query and document tokens. We propose AligneR, a novel multi-vector retrieval model that learns sparsified pairwise alignments between query and document tokens (e.g. `dog' vs. `puppy') and per-token unary saliences reflecting their relative importance for retrieval. We show that controlling the sparsity of pairwise token alignments often brings significant performance gains. While most factoid questions focusing on a specific part of a document require a smaller number of alignments, others requiring a broader understanding of a document favor a larger number of alignments. Unary saliences, on the other hand, decide whether a token ever needs to be aligned with others for retrieval (e.g. `kind' from `kind of currency is used in new zealand}'). With sparsified unary saliences, we are able to prune a large number of query and document token vectors and improve the efficiency of multi-vector retrieval. We learn the sparse unary saliences with entropy-regularized linear programming, which outperforms other methods to achieve sparsity. In a zero-shot setting, AligneR scores 51.1 points nDCG@10, achieving a new retriever-only state-of-the-art on 13 tasks in the BEIR benchmark. In addition, adapting pairwise alignments with a few examples (<= 8) further improves the performance up to 15.7 points nDCG@10 for argument retrieval tasks. The unary saliences of AligneR helps us to keep only 20% of the document token representations with minimal performance loss. We further show that our model often produces interpretable alignments and significantly improves its performance when initialized from larger language models.

Tao Lei, Ran Tian, Jasmijn Bastings et al. · deepmind

In this work, we explore whether modeling recurrence into the Transformer architecture can both be beneficial and efficient, by building an extremely simple recurrent module into the Transformer. We compare our model to baselines following the training and evaluation recipe of BERT. Our results confirm that recurrence can indeed improve Transformer models by a consistent margin, without requiring low-level performance optimizations, and while keeping the number of parameters constant. For example, our base model achieves an absolute improvement of 2.1 points averaged across 10 tasks and also demonstrates increased stability in fine-tuning over a range of learning rates.

Tao Lei, Yetong Xu, Hailong Ning et al.

Popular Transformer networks have been successfully applied to remote sensing (RS) image change detection (CD) identifications and achieve better results than most convolutional neural networks (CNNs), but they still suffer from two main problems. First, the computational complexity of the Transformer grows quadratically with the increase of image spatial resolution, which is unfavorable to very high-resolution (VHR) RS images. Second, these popular Transformer networks tend to ignore the importance of fine-grained features, which results in poor edge integrity and internal tightness for largely changed objects and leads to the loss of small changed objects. To address the above issues, this Letter proposes a Lightweight Structure-aware Transformer (LSAT) network for RS image CD. The proposed LSAT has two advantages. First, a Cross-dimension Interactive Self-attention (CISA) module with linear complexity is designed to replace the vanilla self-attention in visual Transformer, which effectively reduces the computational complexity while improving the feature representation ability of the proposed LSAT. Second, a Structure-aware Enhancement Module (SAEM) is designed to enhance difference features and edge detail information, which can achieve double enhancement by difference refinement and detail aggregation so as to obtain fine-grained features of bi-temporal RS images. Experimental results show that the proposed LSAT achieves significant improvement in detection accuracy and offers a better tradeoff between accuracy and computational costs than most state-of-the-art CD methods for VHR RS images.

Xiaogang Du, Jiawei Zhang, Tongfei Liu et al.

In medical image segmentation tasks, the domain gap caused by the difference in data collection between training and testing data seriously hinders the deployment of pre-trained models in clinical practice. Continual Test-Time Adaptation (CTTA) aims to enable pre-trained models to adapt to continuously changing unlabeled domains, providing an effective approach to solving this problem. However, existing CTTA methods often rely on unreliable supervisory signals, igniting a self-reinforcing cycle of error accumulation that culminates in catastrophic performance degradation. To overcome these challenges, we propose a CTTA via Semantic-Prompt-Enhanced Graph Clustering (SPEGC) for medical image segmentation. First, we design a semantic prompt feature enhancement mechanism that utilizes decoupled commonality and heterogeneity prompt pools to inject global contextual information into local features, alleviating their susceptibility to noise interference under domain shift. Second, based on these enhanced features, we design a differentiable graph clustering solver. This solver reframes global edge sparsification as an optimal transport problem, allowing it to distill a raw similarity matrix into a refined and high-order structural representation in an end-to-end manner. Finally, this robust structural representation is used to guide model adaptation, ensuring predictions are consistent at a cluster-level and dynamically adjusting decision boundaries. Extensive experiments demonstrate that SPEGC outperforms other state-of-the-art CTTA methods on two medical image segmentation benchmarks. The source code is available at https://github.com/Jwei-Z/SPEGC-for-MIS.

Dewen Zeng, Nan Du, Tao Wang et al.

Overparameterized large-scale language models have impressive generalization performance of in-context few-shot learning. However, most language models allocate the same amount of parameters or computation to each token, disregarding the complexity or importance of the input data. We argue that in language model pretraining, a variable amount of computation should be assigned to different tokens, and this can be efficiently achieved via a simple routing mechanism. Different from conventional early stopping techniques where tokens can early exit at only early layers, we propose a more general method that dynamically skips the execution of a layer (or module) for any input token with a binary router. In our extensive evaluation across 24 NLP tasks, we demonstrate that the proposed method can significantly improve the 1-shot performance compared to other competitive baselines only at mild extra cost for inference.

Tom Gunter, Zirui Wang, Chong Wang et al.

We present foundation language models developed to power Apple Intelligence features, including a ~3 billion parameter model designed to run efficiently on devices and a large server-based language model designed for Private Cloud Compute. These models are designed to perform a wide range of tasks efficiently, accurately, and responsibly. This report describes the model architecture, the data used to train the model, the training process, how the models are optimized for inference, and the evaluation results. We highlight our focus on Responsible AI and how the principles are applied throughout the model development.

Xiaogang Du, Dongxin Gu, Tao Lei et al.

In recent years, encoder-decoder networks have focused on expanding receptive fields and incorporating multi-scale context to capture global features for objects of varying sizes. However, as networks deepen, they often discard fine spatial details, impairing precise object localization. Additionally, conventional decoders' use of interpolation for upsampling leads to a loss of global context, diminishing edge segmentation accuracy. To address the above problems, we propose a novel parallel multi-resolution encoder-decoder network, namely PMR-Net for short. First, we design a parallel multi-resolution encoder and a multi-resolution context encoder. The parallel multi-resolution encoder can extract and fuse multi-scale fine-grained local features in parallel for input images with different resolutions. The multi-resolution context encoder fuses the global context semantic features of different receptive fields from different encoder branches to maintain effectively the integrity of global information. Secondly, we design a parallel multi-resolution decoder symmetrical to the structure of parallel multi-resolution encoder. The decoder can continuously supplement the global context features of low-resolution branches to the feature maps of high-resolution branches, and effectively solve the problem of global context feature loss caused by upsampling operation in the decoding process. Extensive experiment results demonstrate that our proposed PMR-Net can achieve more accurate segmentation results than state-of-the-art methods on five public available datasets. Moreover, PMR-Net is also a flexible network framework, which can meet the requirements of different scenarios by adjusting the number of network layers and the number of parallel encoder-decoder branches.

Bailin Wang, Dan Friedman, Tao Lei et al.

Block-wise sparse attention offers significant efficiency gains for long-context modeling, yet existing methods often suffer from low selection fidelity and cumulative contextual loss by completely discarding unselected blocks. To address these limitations, we introduce Sparse Plus Linear Attention (SPLA), a framework that utilizes a selection metric derived from second-order Taylor expansions to accurately identify relevant blocks for exact attention. Instead of discarding the remaining "long tail," SPLA compresses unselected blocks into a compact recurrent state via a residual linear attention (RLA) module. Crucially, to avoid IO overhead, we derive an optimized subtraction-based formulation for RLA -- calculating the residual as the difference between global and selected linear attention -- ensuring that unselected blocks are never explicitly accessed during inference. Our experiments demonstrate that SPLA closes the performance gap in continual pretraining, surpassing dense attention models on long-context benchmarks like RULER while maintaining competitive general knowledge and reasoning capabilities.

Zeyu Cheng, Tongfei Liu, Tao Lei et al.

Depth information is crucial for autonomous driving and intelligent robot navigation. The simplicity and flexibility of self-supervised monocular depth estimation are conducive to its role in these fields. However, most existing monocular depth estimation models consume many computing resources. Although some methods have reduced the model's size and improved computing efficiency, the performance deteriorates, seriously hindering the real-world deployment of self-supervised monocular depth estimation models in the real world. To address this problem, we proposed a real-time self-supervised monocular depth estimation method and implemented it in the real world. It is called RTS-Mono, which is a lightweight and efficient encoder-decoder architecture. The encoder is based on Lite-Encoder, and the decoder is designed with a multi-scale sparse fusion framework to minimize redundancy, ensure performance, and improve inference speed. RTS-Mono achieved state-of-the-art (SoTA) performance in high and low resolutions with extremely low parameter counts (3 M) in experiments based on the KITTI dataset. Compared with lightweight methods, RTS-Mono improved Abs Rel and Sq Rel by 5.6% and 9.8% at low resolution and improved Sq Rel and RMSE by 6.1% and 1.9% at high resolution. In real-world deployment experiments, RTS-Mono has extremely high accuracy and can perform real-time inference on Nvidia Jetson Orin at a speed of 49 FPS. Source code is available at https://github.com/ZYCheng777/RTS-Mono.

Darsh J Shah, Lili Yu, Tao Lei et al.

We introduce \emph{Nutri-bullets}, a multi-document summarization task for health and nutrition. First, we present two datasets of food and health summaries from multiple scientific studies. Furthermore, we propose a novel \emph{extract-compose} model to solve the problem in the regime of limited parallel data. We explicitly select key spans from several abstracts using a policy network, followed by composing the selected spans to present a summary via a task specific language model. Compared to state-of-the-art methods, our approach leads to more faithful, relevant and diverse summarization -- properties imperative to this application. For instance, on the BreastCancer dataset our approach gets a more than 50\% improvement on relevance and faithfulness.\footnote{Our code and data is available at \url{https://github.com/darsh10/Nutribullets.}}

Tao Lei, Xiaohong Jia, Tongliang Liu et al.

Morphological reconstruction (MR) is often employed by seeded image segmentation algorithms such as watershed transform and power watershed as it is able to filter seeds (regional minima) to reduce over-segmentation. However, MR might mistakenly filter meaningful seeds that are required for generating accurate segmentation and it is also sensitive to the scale because a single-scale structuring element is employed. In this paper, a novel adaptive morphological reconstruction (AMR) operation is proposed that has three advantages. Firstly, AMR can adaptively filter useless seeds while preserving meaningful ones. Secondly, AMR is insensitive to the scale of structuring elements because multiscale structuring elements are employed. Finally, AMR has two attractive properties: monotonic increasingness and convergence that help seeded segmentation algorithms to achieve a hierarchical segmentation. Experiments clearly demonstrate that AMR is useful for improving algorithms of seeded image segmentation and seed-based spectral segmentation. Compared to several state-of-the-art algorithms, the proposed algorithms provide better segmentation results requiring less computing time. Source code is available at https://github.com/SUST-reynole/AMR.

Bairu Hou, Qibin Chen, Jianyu Wang et al.

With the rapid scaling of large language models (LLMs), structured pruning has become a widely used technique to learn efficient, smaller models from larger ones, delivering superior performance compared to training similarly sized models from scratch. In this paper, we move beyond the traditional static pruning approach of determining a fixed pruning mask for a model, and propose a dynamic approach to structured pruning. In our method, the pruning mask is input-dependent and adapts dynamically based on the information described in a user instruction. Our approach, termed "instruction-following pruning", introduces a sparse mask predictor that takes the user instruction as input and dynamically selects the most relevant model parameters for the given task. To identify and activate effective parameters, we jointly optimize the sparse mask predictor and the LLM, leveraging both instruction-following data and the pre-training corpus. Experimental results demonstrate the effectiveness of our approach on a wide range of evaluation benchmarks. For example, our 3B activated model improves over the 3B dense model by 5-8 points of absolute margin on domains such as math and coding, and rivals the performance of a 9B model.

Ethan Li, Anders Boesen Lindbo Larsen, Chen Zhang et al. · apple-ml, cmu

We introduce two multilingual, multimodal foundation language models that power Apple Intelligence features across Apple devices and services: i a 3B-parameter on-device model optimized for Apple silicon through architectural innovations such as KV-cache sharing and 2-bit quantization-aware training; and ii a scalable server model built on a novel Parallel-Track Mixture-of-Experts PT-MoE transformer that combines track parallelism, mixture-of-experts sparse computation, and interleaved global-local attention to deliver high quality with competitive cost on Apple's Private Cloud Compute platform. Both models are trained on large-scale multilingual and multimodal datasets sourced via responsible web crawling, licensed corpora, and high-quality synthetic data, then further refined with supervised fine-tuning and reinforcement learning on a new asynchronous platform. The resulting models support several additional languages while understanding images and executing tool calls. In public benchmarks and human evaluations, both the server model and the on-device model match or surpass comparably sized open baselines. A new Swift-centric Foundation Models framework exposes guided generation, constrained tool calling, and LoRA adapter fine-tuning, allowing developers to integrate these capabilities with a few lines of code. The latest advancements in Apple Intelligence models are grounded in our Responsible AI approach with safeguards like content filtering and locale-specific evaluation, as well as our commitment to protecting our users' privacy with innovations like Private Cloud Compute.

Mark Lee, Tom Gunter, Chang Lan et al.

We design and implement AXLearn, a production deep learning system that facilitates scalable and high-performance training of large deep learning models. Compared to other state-of-the-art deep learning systems, AXLearn has a unique focus on modularity and support for heterogeneous hardware infrastructure. AXLearn's internal interfaces between software components follow strict encapsulation, allowing different components to be assembled to facilitate rapid model development and experimentation on heterogeneous compute infrastructure. We introduce a novel method of quantifying modularity via Lines-of-Code (LoC)-complexity, which demonstrates how our system maintains constant complexity as we scale the components in the system, compared to linear or quadratic complexity in other systems. This allows integrating features such as Rotary Position Embeddings (RoPE) into AXLearn across hundred of modules with just 10 lines of code, compared to hundreds as required in other systems. At the same time, AXLearn maintains equivalent performance compared to state-of-the-art training systems. Finally, we share our experience in the development and operation of AXLearn.

Ke Li, Ting Wang, Di Wang et al.

Remote sensing visual grounding (RSVG) aims to localize objects in remote sensing imagery according to natural language expressions. Previous methods typically rely on sentence-level vision-language alignment, which struggles to exploit fine-grained linguistic cues, such as \textit{spatial relations} and \textit{object attributes}, that are crucial for distinguishing objects with similar characteristics. Importantly, these cues play distinct roles across different grounding stages and should be leveraged accordingly to provide more explicit guidance. In this work, we propose \textbf{ProVG}, a novel RSVG framework that improves localization accuracy by decoupling language expressions into global context, spatial relations, and object attributes. To integrate these linguistic cues, ProVG employs a simple yet effective progressive cross-modal modulator, which dynamically modulates visual attention through a \textit{survey-locate-verify} scheme, enabling coarse-to-fine vision-language alignment. In addition, ProVG incorporates a cross-scale fusion module to mitigate the large-scale variations in remote sensing imagery, along with a language-guided calibration decoder to refine cross-modal alignment during prediction. A unified multi-task head further enables ProVG to support both referring expression comprehension and segmentation tasks. Extensive experiments on two benchmarks, \textit{i.e.}, RRSIS-D and RISBench, demonstrate that ProVG consistently outperforms existing methods, achieving new state-of-the-art performance.

Long Shi, Yunshan Ye, Wenjie Wang et al.

Information theory has inspired numerous advancements in multi-view learning. Most multi-view methods incorporating information-theoretic principles rely an assumption called multi-view redundancy which states that common information between views is necessary and sufficient for down-stream tasks. This assumption emphasizes the importance of common information for prediction, but inherently ignores the potential of unique information in each view that could be predictive to the task. In this paper, we propose a comprehensive information-theoretic multi-view learning framework named CIML, which discards the assumption of multi-view redundancy. Specifically, CIML considers the potential predictive capabilities of both common and unique information based on information theory. First, the common representation learning maximizes Gacs-Korner common information to extract shared features and then compresses this information to learn task-relevant representations based on the Information Bottleneck (IB). For unique representation learning, IB is employed to achieve the most compressed unique representation for each view while simultaneously minimizing the mutual information between unique and common representations, as well as among different unique representations. Importantly, we theoretically prove that the learned joint representation is predictively sufficient for the downstream task. Extensive experimental results have demonstrated the superiority of our model over several state-of-art methods. The code is released on CIML.

Yixiao Li, Xianzhi Du, Ajay Jaiswal et al.

Recent advancements in large language models have intensified the need for efficient and deployable models within limited inference budgets. Structured pruning pipelines have shown promise in token efficiency compared to training target-size models from scratch. In this paper, we advocate incorporating enlarged model pretraining, which is often ignored in previous works, into pruning. We study the enlarge-and-prune pipeline as an integrated system to address two critical questions: whether it is worth pretraining an enlarged model even when the model is never deployed, and how to optimize the entire pipeline for better pruned models. We propose an integrated enlarge-and-prune pipeline, which combines enlarge model training, pruning, and recovery under a single cosine annealing learning rate schedule. This approach is further complemented by a novel iterative structured pruning method for gradual parameter removal. The proposed method helps to mitigate the knowledge loss caused by the rising learning rate in naive enlarge-and-prune pipelines and enable effective redistribution of model capacity among surviving neurons, facilitating smooth compression and enhanced performance. We conduct comprehensive experiments on compressing 2.8B models to 1.3B with up to 2T tokens in pretraining. It demonstrates the integrated approach not only provides insights into the token efficiency of enlarged model pretraining but also achieves superior performance of pruned models.

Hailong Ning, Siying Wang, Tao Lei et al.

Remote Sensing Image-Text Retrieval (RSITR) plays a critical role in geographic information interpretation, disaster monitoring, and urban planning by establishing semantic associations between image and textual descriptions. Existing Parameter-Efficient Fine-Tuning (PEFT) methods for Vision-and-Language Pre-training (VLP) models typically adopt symmetric adapter structures for exploring cross-modal correlations. However, the strong discriminative nature of text modality may dominate the optimization process and inhibits image representation learning. The nonnegligible imbalanced cross-modal optimization remains a bottleneck to enhancing the model performance. To address this issue, this study proposes a Representation Discrepancy Bridging (RDB) method for the RSITR task. On the one hand, a Cross-Modal Asymmetric Adapter (CMAA) is designed to enable modality-specific optimization and improve feature alignment. The CMAA comprises a Visual Enhancement Adapter (VEA) and a Text Semantic Adapter (TSA). VEA mines fine-grained image features by Differential Attention (DA) mechanism, while TSA identifies key textual semantics through Hierarchical Attention (HA) mechanism. On the other hand, this study extends the traditional single-task retrieval framework to a dual-task optimization framework and develops a Dual-Task Consistency Loss (DTCL). The DTCL improves cross-modal alignment robustness through an adaptive weighted combination of cross-modal, classification, and exponential moving average consistency constraints. Experiments on RSICD and RSITMD datasets show that the proposed RDB method achieves a 6%-11% improvement in mR metrics compared to state-of-the-art PEFT methods and a 1.15%-2% improvement over the full fine-tuned GeoRSCLIP model.

Xi He, Feiyu Du, Xiaohan Yu et al.

The scarcity of labelled data is specifically an urgent challenge in the field of quantum machine learning (QML). Two transfer fusion frameworks are proposed in this paper to predict the labels of a target domain data by aligning its distribution to a different but related labelled source domain on quantum devices. The frameworks fuses the quantum data from two different, but related domains through a quantum information infusion channel. The predicting tasks in the target domain can be achieved with quantum advantages by post-processing quantum measurement results. One framework, the quantum basic linear algebra subroutines (QBLAS) based implementation, can theoretically achieve the procedure of transfer fusion with quadratic speedup on a universal quantum computer. In addition, the other framework, a hardware-scalable architecture, is implemented on the noisy intermediate-scale quantum (NISQ) devices through a variational hybrid quantum-classical procedure. Numerical experiments on the synthetic and handwritten digits datasets demonstrate that the variatioinal transfer fusion (TF) framework can reach state-of-the-art (SOTA) quantum DA method performance.

Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier et al.

In this work, we discuss building performant Multimodal Large Language Models (MLLMs). In particular, we study the importance of various architecture components and data choices. Through careful and comprehensive ablations of the image encoder, the vision language connector, and various pre-training data choices, we identified several crucial design lessons. For example, we demonstrate that for large-scale multimodal pre-training using a careful mix of image-caption, interleaved image-text, and text-only data is crucial for achieving state-of-the-art (SOTA) few-shot results across multiple benchmarks, compared to other published pre-training results. Further, we show that the image encoder together with image resolution and the image token count has substantial impact, while the vision-language connector design is of comparatively negligible importance. By scaling up the presented recipe, we build MM1, a family of multimodal models up to 30B parameters, including both dense models and mixture-of-experts (MoE) variants, that are SOTA in pre-training metrics and achieve competitive performance after supervised fine-tuning on a range of established multimodal benchmarks. Thanks to large-scale pre-training, MM1 enjoys appealing properties such as enhanced in-context learning, and multi-image reasoning, enabling few-shot chain-of-thought prompting.

Yanqi Zhou, Tao Lei, Hanxiao Liu et al.

Sparsely-activated Mixture-of-experts (MoE) models allow the number of parameters to greatly increase while keeping the amount of computation for a given token or a given sample unchanged. However, a poor expert routing strategy (e.g. one resulting in load imbalance) can cause certain experts to be under-trained, leading to an expert being under or over-specialized. Prior work allocates a fixed number of experts to each token using a top-k function regardless of the relative importance of different tokens. To address this, we propose a heterogeneous mixture-of-experts employing an expert choice method. Instead of letting tokens select the top-k experts, we have experts selecting the top-k tokens. As a result, each token can be routed to a variable number of experts and each expert can have a fixed bucket size. We systematically study pre-training speedups using the same computational resources of the Switch Transformer top-1 and GShard top-2 gating of prior work and find that our method improves training convergence time by more than 2x. For the same computational cost, our method demonstrates higher performance in fine-tuning 11 selected tasks in the GLUE and SuperGLUE benchmarks. For a smaller activation cost, our method outperforms the T5 dense model in 7 out of the 11 tasks.

Jing Pan, Tao Lei, Kwangyoun Kim et al.

The Transformer architecture has been well adopted as a dominant architecture in most sequence transduction tasks including automatic speech recognition (ASR), since its attention mechanism excels in capturing long-range dependencies. While models built solely upon attention can be better parallelized than regular RNN, a novel network architecture, SRU++, was recently proposed. By combining the fast recurrence and attention mechanism, SRU++ exhibits strong capability in sequence modeling and achieves near-state-of-the-art results in various language modeling and machine translation tasks with improved compute efficiency. In this work, we present the advantages of applying SRU++ in ASR tasks by comparing with Conformer across multiple ASR benchmarks and study how the benefits can be generalized to long-form speech inputs. On the popular LibriSpeech benchmark, our SRU++ model achieves 2.0% / 4.7% WER on test-clean / test-other, showing competitive performances compared with the state-of-the-art Conformer encoder under the same set-up. Specifically, SRU++ can surpass Conformer on long-form speech input with a large margin, based on our analysis.

Xianyuan Liu, Shuo Zhou, Tao Lei et al.

Unsupervised Domain Adaptation (UDA) can transfer knowledge from labeled source data to unlabeled target data of the same categories. However, UDA for first-person action recognition is an under-explored problem, with lack of datasets and limited consideration of first-person video characteristics. This paper focuses on addressing this problem. Firstly, we propose two small-scale first-person video domain adaptation datasets: ADL$_{small}$ and GTEA-KITCHEN. Secondly, we introduce channel-temporal attention blocks to capture the channel-wise and temporal-wise relationships and model their inter-dependencies important to first-person vision. Finally, we propose a Channel-Temporal Attention Network (CTAN) to integrate these blocks into existing architectures. CTAN outperforms baselines on the two proposed datasets and one existing dataset EPIC$_{cvpr20}$.

Xianyuan Liu, Raivo Koot, Shuo Zhou et al.

This report describes the technical details of our submission to the EPIC-Kitchens 2021 Unsupervised Domain Adaptation Challenge for Action Recognition. The EPIC-Kitchens dataset is more difficult than other video domain adaptation datasets due to multi-tasks with more modalities. Firstly, to participate in the challenge, we employ a transformer to capture the spatial information from each modality. Secondly, we employ a temporal attention module to model temporal-wise inter-dependency. Thirdly, we employ the adversarial domain adaptation network to learn the general features between labeled source and unlabeled target domain. Finally, we incorporate multiple modalities to improve the performance by a three-stream network with late fusion. Our network achieves the comparable performance with the state-of-the-art baseline T$A^3$N and outperforms the baseline on top-1 accuracy for verb class and top-5 accuracies for all three tasks which are verb, noun and action. Under the team name xy9, our submission achieved 5th place in terms of top-1 accuracy for verb class and all top-5 accuracies.

Darsh J Shah, Lili Yu, Tao Lei et al.

We present a method for generating comparative summaries that highlights similarities and contradictions in input documents. The key challenge in creating such summaries is the lack of large parallel training data required for training typical summarization systems. To this end, we introduce a hybrid generation approach inspired by traditional concept-to-text systems. To enable accurate comparison between different sources, the model first learns to extract pertinent relations from input documents. The content planning component uses deterministic operators to aggregate these relations after identifying a subset for inclusion into a summary. The surface realization component lexicalizes this information using a text-infilling language model. By separately modeling content selection and realization, we can effectively train them with limited annotations. We implemented and tested the model in the domain of nutrition and health -- rife with inconsistencies. Compared to conventional methods, our framework leads to more faithful, relevant and aggregation-sensitive summarization -- while being equally fluent.

Tao Lei

Large language models have become increasingly difficult to train because of the growing computation time and cost. In this work, we present SRU++, a highly-efficient architecture that combines fast recurrence and attention for sequence modeling. SRU++ exhibits strong modeling capacity and training efficiency. On standard language modeling tasks such as Enwik8, Wiki-103 and Billion Word datasets, our model obtains better bits-per-character and perplexity while using 3x-10x less training cost compared to top-performing Transformer models. For instance, our model achieves a state-of-the-art result on the Enwik8 dataset using 1.6 days of training on an 8-GPU machine. We further demonstrate that SRU++ requires minimal attention for near state-of-the-art performance. Our results suggest jointly leveraging fast recurrence with little attention as a promising direction for accelerating model training and inference.

Risheng Wang, Tao Lei, Ruixia Cui et al.

Deep learning has been widely used for medical image segmentation and a large number of papers has been presented recording the success of deep learning in the field. In this paper, we present a comprehensive thematic survey on medical image segmentation using deep learning techniques. This paper makes two original contributions. Firstly, compared to traditional surveys that directly divide literatures of deep learning on medical image segmentation into many groups and introduce literatures in detail for each group, we classify currently popular literatures according to a multi-level structure from coarse to fine. Secondly, this paper focuses on supervised and weakly supervised learning approaches, without including unsupervised approaches since they have been introduced in many old surveys and they are not popular currently. For supervised learning approaches, we analyze literatures in three aspects: the selection of backbone networks, the design of network blocks, and the improvement of loss functions. For weakly supervised learning approaches, we investigate literature according to data augmentation, transfer learning, and interactive segmentation, separately. Compared to existing surveys, this survey classifies the literatures very differently from before and is more convenient for readers to understand the relevant rationale and will guide them to think of appropriate improvements in medical image segmentation based on deep learning approaches.

Alexander Lin, Jeremy Wohlwend, Howard Chen et al.

The performance of autoregressive models on natural language generation tasks has dramatically improved due to the adoption of deep, self-attentive architectures. However, these gains have come at the cost of hindering inference speed, making state-of-the-art models cumbersome to deploy in real-world, time-sensitive settings. We develop a compression technique for autoregressive models that is driven by an imitation learning perspective on knowledge distillation. The algorithm is designed to address the exposure bias problem. On prototypical language generation tasks such as translation and summarization, our method consistently outperforms other distillation algorithms, such as sequence-level knowledge distillation. Student models trained with our method attain 1.4 to 4.8 BLEU/ROUGE points higher than those trained from scratch, while increasing inference speed by up to 14 times in comparison to the teacher model.

Kyle Swanson, Lili Yu, Tao Lei

Selecting input features of top relevance has become a popular method for building self-explaining models. In this work, we extend this selective rationalization approach to text matching, where the goal is to jointly select and align text pieces, such as tokens or sentences, as a justification for the downstream prediction. Our approach employs optimal transport (OT) to find a minimal cost alignment between the inputs. However, directly applying OT often produces dense and therefore uninterpretable alignments. To overcome this limitation, we introduce novel constrained variants of the OT problem that result in highly sparse alignments with controllable sparsity. Our model is end-to-end differentiable using the Sinkhorn algorithm for OT and can be trained without any alignment annotations. We evaluate our model on the StackExchange, MultiNews, e-SNLI, and MultiRC datasets. Our model achieves very sparse rationale selections with high fidelity while preserving prediction accuracy compared to strong attention baseline models.

Jing Pan, Joshua Shapiro, Jeremy Wohlwend et al.

In this paper we present state-of-the-art (SOTA) performance on the LibriSpeech corpus with two novel neural network architectures, a multistream CNN for acoustic modeling and a self-attentive simple recurrent unit (SRU) for language modeling. In the hybrid ASR framework, the multistream CNN acoustic model processes an input of speech frames in multiple parallel pipelines where each stream has a unique dilation rate for diversity. Trained with the SpecAugment data augmentation method, it achieves relative word error rate (WER) improvements of 4% on test-clean and 14% on test-other. We further improve the performance via N-best rescoring using a 24-layer self-attentive SRU language model, achieving WERs of 1.75% on test-clean and 4.46% on test-other.

Lili Yu, Howard Chen, Sida Wang et al.

We study the potential for interaction in natural language classification. We add a limited form of interaction for intent classification, where users provide an initial query using natural language, and the system asks for additional information using binary or multi-choice questions. At each turn, our system decides between asking the most informative question or making the final classification prediction.The simplicity of the model allows for bootstrapping of the system without interaction data, instead relying on simple crowdsourcing tasks. We evaluate our approach on two domains, showing the benefit of interaction and the advantage of learning to balance between asking additional questions and making the final prediction.

Jeremy Wohlwend, Ethan R. Elenberg, Samuel Altschul et al.

Traditional text classifiers are limited to predicting over a fixed set of labels. However, in many real-world applications the label set is frequently changing. For example, in intent classification, new intents may be added over time while others are removed. We propose to address the problem of dynamic text classification by replacing the traditional, fixed-size output layer with a learned, semantically meaningful metric space. Here the distances between textual inputs are optimized to perform nearest-neighbor classification across overlapping label sets. Changing the label set does not involve removing parameters, but rather simply adding or removing support points in the metric space. Then the learned metric can be fine-tuned with only a few additional training examples. We demonstrate that this simple strategy is robust to changes in the label space. Furthermore, our results show that learning a non-Euclidean metric can improve performance in the low data regime, suggesting that further work on metric spaces may benefit low-resource research.

Ziheng Wang, Jeremy Wohlwend, Tao Lei

Large language models have recently achieved state of the art performance across a wide variety of natural language tasks. Meanwhile, the size of these models and their latency have significantly increased, which makes their usage costly, and raises an interesting question: do language models need to be large? We study this question through the lens of model compression. We present a generic, structured pruning approach by parameterizing each weight matrix using its low-rank factorization, and adaptively removing rank-1 components during training. On language modeling tasks, our structured approach outperforms other unstructured and block-structured pruning baselines at various compression levels, while achieving significant speedups during both training and inference. We also demonstrate that our method can be applied to pruning adaptive word embeddings in large language models, and to pruning the BERT model on several downstream fine-tuning classification benchmarks.

Kyle Swanson, Lili Yu, Christopher Fox et al.

Response suggestion is an important task for building human-computer conversation systems. Recent approaches to conversation modeling have introduced new model architectures with impressive results, but relatively little attention has been paid to whether these models would be practical in a production setting. In this paper, we describe the unique challenges of building a production retrieval-based conversation system, which selects outputs from a whitelist of candidate responses. To address these challenges, we propose a dual encoder architecture which performs rapid inference and scales well with the size of the whitelist. We also introduce and compare two methods for generating whitelists, and we carry out a comprehensive analysis of the model and whitelists. Experimental results on a large, proprietary help desk chat dataset, including both offline metrics and a human evaluation, indicate production-quality performance and illustrate key lessons about conversation modeling in practice.

Darsh J Shah, Tao Lei, Alessandro Moschitti et al.

We address the problem of detecting duplicate questions in forums, which is an important step towards automating the process of answering new questions. As finding and annotating such potential duplicates manually is very tedious and costly, automatic methods based on machine learning are a viable alternative. However, many forums do not have annotated data, i.e., questions labeled by experts as duplicates, and thus a promising solution is to use domain adaptation from another forum that has such annotations. Here we focus on adversarial domain adaptation, deriving important findings about when it performs well and what properties of the domains are important in this regard. Our experiments with StackExchange data show an average improvement of 5.6% over the best baseline across multiple pairs of domains.

Tao Lei, Yu Zhang, Sida I. Wang et al.

Common recurrent neural architectures scale poorly due to the intrinsic difficulty in parallelizing their state computations. In this work, we propose the Simple Recurrent Unit (SRU), a light recurrent unit that balances model capacity and scalability. SRU is designed to provide expressive recurrence, enable highly parallelized implementation, and comes with careful initialization to facilitate training of deep models. We demonstrate the effectiveness of SRU on multiple NLP tasks. SRU achieves 5--9x speed-up over cuDNN-optimized LSTM on classification and question answering datasets, and delivers stronger results than LSTM and convolutional models. We also obtain an average of 0.7 BLEU improvement over the Transformer model on translation by incorporating SRU into the architecture.

Tianxiao Shen, Tao Lei, Regina Barzilay et al.

This paper focuses on style transfer on the basis of non-parallel text. This is an instance of a broad family of problems including machine translation, decipherment, and sentiment modification. The key challenge is to separate the content from other aspects such as style. We assume a shared latent content distribution across different text corpora, and propose a method that leverages refined alignment of latent representations to perform style transfer. The transferred sentences from one style should match example sentences from the other style as a population. We demonstrate the effectiveness of this cross-alignment method on three tasks: sentiment modification, decipherment of word substitution ciphers, and recovery of word order.

Tao Lei, Wengong Jin, Regina Barzilay et al.

The design of neural architectures for structured objects is typically guided by experimental insights rather than a formal process. In this work, we appeal to kernels over combinatorial structures, such as sequences and graphs, to derive appropriate neural operations. We introduce a class of deep recurrent neural operations and formally characterize their associated kernel spaces. Our recurrent modules compare the input to virtual reference objects (cf. filters in CNN) via the kernels. Similar to traditional neural operations, these reference objects are parameterized and directly optimized in end-to-end training. We empirically evaluate the proposed class of neural architectures on standard applications such as language modeling and molecular graph regression, achieving state-of-the-art results across these applications.

Tao Lei, Regina Barzilay, Tommi Jaakkola

Prediction without justification has limited applicability. As a remedy, we learn to extract pieces of input text as justifications -- rationales -- that are tailored to be short and coherent, yet sufficient for making the same prediction. Our approach combines two modular components, generator and encoder, which are trained to operate well together. The generator specifies a distribution over text fragments as candidate rationales and these are passed through the encoder for prediction. Rationales are never given during training. Instead, the model is regularized by desiderata for rationales. We evaluate the approach on multi-aspect sentiment analysis against manually annotated test cases. Our approach outperforms attention-based baseline by a significant margin. We also successfully illustrate the method on the question retrieval task.

Tao Lei, Hrishikesh Joshi, Regina Barzilay et al.

Question answering forums are rapidly growing in size with no effective automated ability to refer to and reuse answers already available for previous posted questions. In this paper, we develop a methodology for finding semantically related questions. The task is difficult since 1) key pieces of information are often buried in extraneous details in the question body and 2) available annotations on similar questions are scarce and fragmented. We design a recurrent and convolutional model (gated convolution) to effectively map questions to their semantic representations. The models are pre-trained within an encoder-decoder framework (from body to title) on the basis of the entire raw corpus, and fine-tuned discriminatively from limited annotations. Our evaluation demonstrates that our model yields substantial gains over a standard IR baseline and various neural network architectures (including CNNs, LSTMs and GRUs).

Tao Lei, Regina Barzilay, Tommi Jaakkola

The success of deep learning often derives from well-chosen operational building blocks. In this work, we revise the temporal convolution operation in CNNs to better adapt it to text processing. Instead of concatenating word representations, we appeal to tensor algebra and use low-rank n-gram tensors to directly exploit interactions between words already at the convolution stage. Moreover, we extend the n-gram convolution to non-consecutive words to recognize patterns with intervening words. Through a combination of low-rank tensors, and pattern weighting, we can efficiently evaluate the resulting convolution operation via dynamic programming. We test the resulting architecture on standard sentiment classification and news categorization tasks. Our model achieves state-of-the-art performance both in terms of accuracy and training speed. For instance, we obtain 51.2% accuracy on the fine-grained sentiment classification task.