A common concern when a policymaker draws causal inferences from and makes decisions based on observational data is that the measured covariates are insufficiently rich to account for all sources of confounding, i.e., the standard no confoundedness assumption fails to hold. The recently proposed proximal causal inference framework shows that proxy variables can be leveraged to identify causal effects and therefore facilitate decision-making. Building upon this line of work, we propose a novel optimal individualized treatment regime based on so-called outcome-inducing and treatment-inducing confounding bridges. We then show that the value function of this new optimal treatment regime is superior to that of existing ones in the literature. Theoretical guarantees, including identification, superiority, and excess value bound of the estimated regime, are established. Furthermore, we demonstrate the proposed optimal regime via numerical experiments and a real data application.
Long document retrieval aims to fetch query-relevant documents from a large-scale collection, where knowledge distillation has become de facto to improve a retriever by mimicking a heterogeneous yet powerful cross-encoder. However, in contrast to passages or sentences, retrieval on long documents suffers from the scope hypothesis that a long document may cover multiple topics. This maximizes their structure heterogeneity and poses a granular-mismatch issue, leading to an inferior distillation efficacy. In this work, we propose a new learning framework, fine-grained distillation (FGD), for long-document retrievers. While preserving the conventional dense retrieval paradigm, it first produces global-consistent representations crossing different fine granularity and then applies multi-granular aligned distillation merely during training. In experiments, we evaluate our framework on two long-document retrieval benchmarks, which show state-of-the-art performance.
To improve the performance of the dual-encoder retriever, one effective approach is knowledge distillation from the cross-encoder ranker. Existing works construct the candidate passages following the supervised learning setting where a query is paired with a positive passage and a batch of negatives. However, through empirical observation, we find that even the hard negatives from advanced methods are still too trivial for the teacher to distinguish, preventing the teacher from transferring abundant dark knowledge to the student through its soft label. To alleviate this issue, we propose ADAM, a knowledge distillation framework that can better transfer the dark knowledge held in the teacher with Adaptive Dark exAMples. Different from previous works that only rely on one positive and hard negatives as candidate passages, we create dark examples that all have moderate relevance to the query through mixing-up and masking in discrete space. Furthermore, as the quality of knowledge held in different training instances varies as measured by the teacher's confidence score, we propose a self-paced distillation strategy that adaptively concentrates on a subset of high-quality instances to conduct our dark-example-based knowledge distillation to help the student learn better. We conduct experiments on two widely-used benchmarks and verify the effectiveness of our method.
With the success of the prompt-tuning paradigm in Natural Language Processing (NLP), various prompt templates have been proposed to further stimulate specific knowledge for serving downstream tasks, e.g., machine translation, text generation, relation extraction, and so on. Existing prompt templates are mainly shared among all training samples with the information of task description. However, training samples are quite diverse. The sharing task description is unable to stimulate the unique task-related information in each training sample, especially for tasks with the finite-label space. To exploit the unique task-related information, we imitate the human decision process which aims to find the contrastive attributes between the objective factual and their potential counterfactuals. Thus, we propose the \textbf{C}ounterfactual \textbf{C}ontrastive \textbf{Prompt}-Tuning (CCPrompt) approach for many-class classification, e.g., relation classification, topic classification, and entity typing. Compared with simple classification tasks, these tasks have more complex finite-label spaces and are more rigorous for prompts. First of all, we prune the finite label space to construct fact-counterfactual pairs. Then, we exploit the contrastive attributes by projecting training instances onto every fact-counterfactual pair. We further set up global prototypes corresponding with all contrastive attributes for selecting valid contrastive attributes as additional tokens in the prompt template. Finally, a simple Siamese representation learning is employed to enhance the robustness of the model. We conduct experiments on relation classification, topic classification, and entity typing tasks in both fully supervised setting and few-shot setting. The results indicate that our model outperforms former baselines.
The high-content image-based assay is commonly leveraged for identifying the phenotypic impact of genetic perturbations in biology field. However, a persistent issue remains unsolved during experiments: the interferential technical noise caused by systematic errors (e.g., temperature, reagent concentration, and well location) is always mixed up with the real biological signals, leading to misinterpretation of any conclusion drawn. Here, we show a mean teacher based deep learning model (DeepNoise) that can disentangle biological signals from the experimental noise. Specifically, we aim to classify the phenotypic impact of 1,108 different genetic perturbations screened from 125,510 fluorescent microscopy images, which are totally unrecognizable by human eye. We validate our model by participating in the Recursion Cellular Image Classification Challenge, and our proposed method achieves an extremely high classification score (Acc: 99.596%), ranking the 2nd place among 866 participating groups. This promising result indicates the successful separation of biological and technical factors, which might help decrease the cost of treatment development and expedite the drug discovery process.
Deploying machine learning models on mobile devices has gained increasing attention. To tackle the model generalization problem with the limitations of hardware resources on the device, the device model needs to be lightweight by techniques such as model compression from the cloud model. However, the major obstacle to improve the device model generalization is the distribution shift between the data of cloud and device models, since the data distribution on device model often changes over time (e.g., users might have different preferences in recommendation system). Although real-time fine-tuning and distillation method take this situation into account, these methods require on-device training, which are practically infeasible due to the low computational power and a lack of real-time labeled samples on the device. In this paper, we propose a novel task-agnostic framework, named MetaNetwork, for generating adaptive device model parameters from cloud without on-device training. Specifically, our MetaNetwork is deployed on cloud and consists of MetaGenerator and MetaStabilizer modules. The MetaGenerator is designed to learn a mapping function from samples to model parameters, and it can generate and deliver the adaptive parameters to the device based on samples uploaded from the device to the cloud. The MetaStabilizer aims to reduce the oscillation of the MetaGenerator, accelerate the convergence and improve the model performance during both training and inference. We evaluate our method on two tasks with three datasets. Extensive experiments show that MetaNetwork can achieve competitive performances in different modalities.
In large-scale retrieval, the lexicon-weighting paradigm, learning weighted sparse representations in vocabulary space, has shown promising results with high quality and low latency. Despite it deeply exploiting the lexicon-representing capability of pre-trained language models, a crucial gap remains between language modeling and lexicon-weighting retrieval -- the former preferring certain or low-entropy words whereas the latter favoring pivot or high-entropy words -- becoming the main barrier to lexicon-weighting performance for large-scale retrieval. To bridge this gap, we propose a brand-new pre-training framework, lexicon-bottlenecked masked autoencoder (LexMAE), to learn importance-aware lexicon representations. Essentially, we present a lexicon-bottlenecked module between a normal language modeling encoder and a weakened decoder, where a continuous bag-of-words bottleneck is constructed to learn a lexicon-importance distribution in an unsupervised fashion. The pre-trained LexMAE is readily transferred to the lexicon-weighting retrieval via fine-tuning, achieving 42.6\% MRR@10 with 45.83 QPS on a CPU machine for the passage retrieval benchmark, MS-Marco. And LexMAE shows state-of-the-art zero-shot transfer capability on BEIR benchmark with 12 datasets.
Retrieval models based on dense representations in semantic space have become an indispensable branch for first-stage retrieval. These retrievers benefit from surging advances in representation learning towards compressive global sequence-level embeddings. However, they are prone to overlook local salient phrases and entity mentions in texts, which usually play pivot roles in first-stage retrieval. To mitigate this weakness, we propose to make a dense retriever align a well-performing lexicon-aware representation model. The alignment is achieved by weakened knowledge distillations to enlighten the retriever via two aspects -- 1) a lexicon-augmented contrastive objective to challenge the dense encoder and 2) a pair-wise rank-consistent regularization to make dense model's behavior incline to the other. We evaluate our model on three public benchmarks, which shows that with a comparable lexicon-aware retriever as the teacher, our proposed dense one can bring consistent and significant improvements, and even outdo its teacher. In addition, we found our improvement on the dense retriever is complementary to the standard ranker distillation, which can further lift state-of-the-art performance.
RNA structure determination and prediction can promote RNA-targeted drug development and engineerable synthetic elements design. But due to the intrinsic structural flexibility of RNAs, all the three mainstream structure determination methods (X-ray crystallography, NMR, and Cryo-EM) encounter challenges when resolving the RNA structures, which leads to the scarcity of the resolved RNA structures. Computational prediction approaches emerge as complementary to the experimental techniques. However, none of the \textit{de novo} approaches is based on deep learning since too few structures are available. Instead, most of them apply the time-consuming sampling-based strategies, and their performance seems to hit the plateau. In this work, we develop the first end-to-end deep learning approach, E2Efold-3D, to accurately perform the \textit{de novo} RNA structure prediction. Several novel components are proposed to overcome the data scarcity, such as a fully-differentiable end-to-end pipeline, secondary structure-assisted self-distillation, and parameter-efficient backbone formulation. Such designs are validated on the independent, non-overlapping RNA puzzle testing dataset and reach an average sub-4 \AA{} root-mean-square deviation, demonstrating its superior performance compared to state-of-the-art approaches. Interestingly, it also achieves promising results when predicting RNA complex structures, a feat that none of the previous systems could accomplish. When E2Efold-3D is coupled with the experimental techniques, the RNA structure prediction field can be greatly advanced.
This paper focuses on text data augmentation for few-shot NLP tasks. The existing data augmentation algorithms either leverage task-independent heuristic rules (e.g., Synonym Replacement) or fine-tune general-purpose pre-trained language models (e.g., GPT2) using a small training set to produce new synthetic data. Consequently, these methods have trivial task-specific knowledge and are limited to yielding low-quality synthetic data for weak baselines in simple tasks. To combat this issue, we propose the Knowledge Mixture Data Augmentation Model (KnowDA): an encoder-decoder LM pretrained on a mixture of diverse NLP tasks using Knowledge Mixture Training (KoMT). KoMT is a training procedure that reformulates input examples from various heterogeneous NLP tasks into a unified text-to-text format and employs denoising objectives in different granularity to learn to generate partial or complete samples. With the aid of KoMT, KnowDA could combine required task-specific knowledge implicitly from the learned mixture of tasks and quickly grasp the inherent synthesis law of the target task through a few given instances. To the best of our knowledge, we are the first attempt to scale the number of tasks to 100+ in multi-task co-training for data augmentation. Extensive experiments show that i) KnowDA successfully improves the performance of Albert and Deberta by a large margin on the FewGLUE benchmark, outperforming previous state-of-the-art data augmentation baselines; ii) KnowDA could also improve the model performance on the few-shot NER tasks, a held-out task type not included in KoMT.