Syntax and semantics are two key concepts for language understanding. Topic models typically represent the semantics of a text corpus, while removing syntactic information during preprocessing. Without preprocessing, the generated topics become uninterpretable because the syntactic words dominate generated topics. To learn interpretable topics while keeping valuable syntactic information, we propose a novel framework that can simultaneously learn both syntactic and semantic topics from the corpus without requiring any preprocessing. A context network leverages textual dependencies to distinguish between syntactic and semantic words, while a composite VAE topic model learns two sets of topics. We demonstrate on seven datasets that our proposed method effectively captures both syntactic and semantic representations of a corpus while outperforming state-of-the-art neural topic models and statistical topic models in terms of topic quality.
论文检索
输入标题、作者或关键词，从 7,537 篇学术成果中精准定位
The Mixtures-of-Experts (MoE) model is a widespread distributed and integrated learning method for large language models (LLM), which is favored due to its ability to sparsify and expand models efficiently. However, the performance of MoE is limited by load imbalance and high latency of All-to-All communication, along with relatively redundant computation owing to large expert capacity. Load imbalance may result from existing routing policies that consistently tend to select certain experts. The frequent inter-node communication in the All-to-All procedure also significantly prolongs the training time. To alleviate the above performance problems, we propose a novel routing strategy that combines load balance and locality by converting partial inter-node communication to that of intra-node. Notably, we elucidate that there is a minimum threshold for expert capacity, calculated through the maximal angular deviation between the gating weights of the experts and the assigned tokens. We port these modifications on the PanGu-Σ model based on the MindSpore framework with multi-level routing and conduct experiments on Ascend clusters. The experiment results demonstrate that the proposed LocMoE reduces training time per epoch by 12.68% to 22.24% compared to classical routers, such as hash router and switch router, without impacting the model accuracy.
Reframing Spatial Reasoning Evaluation in Language Models: A Real-World Simulation Benchmark for Qualitative Reasoning
PDF ↗Spatial reasoning plays a vital role in both human cognition and machine intelligence, prompting new research into language models' (LMs) capabilities in this regard. However, existing benchmarks reveal shortcomings in evaluating qualitative spatial reasoning (QSR). These benchmarks typically present oversimplified scenarios or unclear natural language descriptions, hindering effective evaluation. We present a novel benchmark for assessing QSR in LMs, which is grounded in realistic 3D simulation data, offering a series of diverse room layouts with various objects and their spatial relationships. This approach provides a more detailed and context-rich narrative for spatial reasoning evaluation, diverging from traditional, toy-task-oriented scenarios. Our benchmark encompasses a broad spectrum of qualitative spatial relationships, including topological, directional, and distance relations. These are presented with different viewing points, varied granularities, and density of relation constraints to mimic real-world complexities. A key contribution is our logic-based consistency-checking tool, which enables the assessment of multiple plausible solutions, aligning with real-world scenarios where spatial relationships are often open to interpretation. Our benchmark evaluation of advanced LMs reveals their strengths and limitations in spatial reasoning. They face difficulties with multi-hop spatial reasoning and interpreting a mix of different view descriptions, pointing to areas for future improvement.
Designing natural language interfaces has historically required collecting supervised data to translate user requests into carefully designed intent representations. This requires enumerating and labeling a long tail of user requests, which is challenging. At the same time, large language models (LLMs) encode knowledge about goals and plans that can help conversational assistants interpret user requests requiring numerous steps to complete. We introduce an approach to handle complex-intent-bearing utterances from a user via a process of hierarchical natural language decomposition and interpretation. Our approach uses a pre-trained language model to decompose a complex utterance into a sequence of simpler natural language steps and interprets each step using the language-to-program model designed for the interface. To test our approach, we collect and release DeCU —a new NL-to-program benchmark to evaluate Decomposition of Complex Utterances. Experiments show that the proposed approach enables the interpretation of complex utterances with almost no complex training data, while outperforming standard few-shot prompting approaches.
In this work, we propose a novel tree-based explanation technique, PEACH (Pretrained-embedding Explanation Across Contextual and Hierarchical Structure), that can explain how text-based documents are classified by using any pretrained contextual embeddings in a tree-based human-interpretable manner. Note that PEACH can adopt any contextual embeddings of the PLMs as a training input for the decision tree. Using the proposed PEACH, we perform a comprehensive analysis of several contextual embeddings on nine different NLP text classification benchmarks. This analysis demonstrates the flexibility of the model by appling several PLM contextual embeddings, its attribute selections, scaling, and clustering methods. Furthermore, we show the utility of explanations by visualising the feature selection and important trend of text classification via human-interpretable word-cloud-based trees, which clearly identify model mistakes and assist in dataset debugging. Besides interpretability, PEACH outperforms or is similar to those from pretrained models. Code and Appendix are in https://github.com/adlnlp/peach.
Memes, initially created for humor and social commentary, have transformed into platforms for offensive online content. Detecting such content is crucial; however, existing deep learning-based meme offensiveness classifiers lack transparency, functioning as opaque black-box systems. While Integrated Gradient and similar input-attribution interpretability methods exist, they often yield inadequate and irrelevant keywords. To bridge this gap, we introduce SEMANTIFY, a novel system featuring a theoretically grounded multi-step filtering process. SEMANTIFY extracts meaningful "tokens" from a predefined vocabulary, generating a pertinent and comprehensive set of interpretable keywords. These extracted keywords reveal the model's awareness of hidden meanings in memes, enhancing transparency. Evaluation of SEMANTIFY using interpretability metrics, including 'leakage-adjusted simulatability,' demonstrates its superiority over various baselines by up to 2.5 points. Human evaluation of 'relatedness' and 'exhaustiveness' of extracted keywords further validates its effectiveness. Additionally, a qualitative analysis of extracted keywords serves as a case study, unveiling model error cases and their reasons. SEMANTIFY contributes to the advancement of more interpretable multimodal systems for meme offensiveness detection, fostering trust for real-world applications.
MMGNN: A Molecular Merged Graph Neural Network for Explainable Solvation Free Energy Prediction
PDF ↗In this paper, we address the challenge of accurately modeling and predicting Gibbs free energy in solute-solvent interactions, a pivotal yet complex aspect in the field of chemical modeling. Traditional approaches, primarily relying on deep learning models, face limitations in capturing the intricate dynamics of these interactions. To overcome these constraints, we introduce a novel framework, molecular modeling graph neural network (MMGNN), which more closely mirrors real-world chemical processes.Specifically, MMGNN explicitly models atomic interactions such as hydrogen bonds by initially forming indiscriminate connections between intermolecular atoms, which are then refined using an attention-based aggregation method, tailoring to specific solute-solvent pairs. To address the challenges of non-interactive or repulsive atomic interactions, MMGNN incorporates interpreters for nodes and edges in the merged graph, enhancing explainability and reducing redundancy. MMGNN stands as the first framework to explicitly align with real chemical processes, providing a more accurate and scientifically sound approach to modeling solute-solvent interactions. The infusion of explainability allows for the extraction of key subgraphs, which are pivotal for further research in solute-solvent dynamics. Extensive experimental validation confirms the efficacy and enhanced explainability of MMGNN.
Delve into Base-Novel Confusion: Redundancy Exploration for Few-Shot Class-Incremental Learning
PDF ↗Few-shot class-incremental learning (FSCIL) aims to acquire knowledge from novel classes with limited samples while retaining information about base classes. Existing methods address catastrophic forgetting and overfitting by freezing the feature extractor during novel-class learning. However, these methods usually tend to cause the confusion between base and novel classes, i.e., classifying novel-class samples into base classes.In this paper, we delve into this phenomenon to study its cause and solution. We first interpret the confusion as the collision between the novel-class and the base-class region in the feature space.Then, we find the collision is caused by the label-irrelevant redundancies within the base-class feature and pixel space. Through qualitative and quantitative experiments, we identify this redundancy as the shortcut in the base-class training, which can be decoupled to alleviate the collision. Based on this analysis, to alleviate the collision between base and novel classes, we propose a method for FSCIL named Redundancy Decoupling and Integration (RDI). RDI first decouples redundancies from base-class space to shrink the intra-base-class feature space. Then, it integrates the redundancies as a dummy class to enlarge the inter-base-class feature space. This process effectively compresses the base-class feature space, creating buffer space for novel classes and alleviating the model's confusion between the base and novel classes. Extensive experiments across benchmark datasets, including CIFAR-100, miniImageNet, and CUB-200-2011 demonstrate that our method achieves state-of-the-art performance.
As the theoretical underpinnings for subspace clustering and classification, subspace-preserving recovery has attracted intensive attention in recent years. However, previous theoretical advances for subspace-preserving recovery only focus on the single-view data and most of them are based on conditions that are only sufficient. In this paper, we propose a necessary and sufficient condition referred to as Atomic Recovery Property (ARP) for multi-view subspace-preserving recovery. To this end, we generalize the atomic norm from single-view data to multi-view data and define the Multi-view Atomic Norm (MAN). Our another contribution is to provide a geometrically more interpretable characterization of ARP with respect to the unit ball of MAN. Based on the proposed multi-view subspace-preserving recovery theory, we also derive novel theoretical results for multi-view subspace clustering and classification, respectively.
Conventional machine learning methods are predominantly designed to predict outcomes based on a single data type. However, practical applications may encompass data of diverse types, such as text, images, and audio. We introduce interpretable tensor fusion (InTense), a multimodal learning method training a neural network to simultaneously learn multiple data representations and their interpretable fusion. InTense can separately capture both linear combinations and multiplicative interactions of the data types, thereby disentangling higher-order interactions from the individual effects of each modality. InTense provides interpretability out of the box by assigning relevance scores to modalities and their associations, respectively. The approach is theoretically grounded and yields meaningful relevance scores on multiple synthetic and real-world datasets. Experiments on four real-world datasets show that InTense outperforms existing state-of-the-art multimodal interpretable approaches in terms of accuracy and interpretability.
Neural networks often assume independence among input data samples, disregarding correlations arising from inherent clustering patterns in real-world datasets (e.g., due to different sites or repeated measurements). Recently, mixed effects neural networks (MENNs) which separate cluster-specific 'random effects' from cluster-invariant 'fixed effects' have been proposed to improve generalization and interpretability for clustered data. However, existing methods only allow for approximate quantification of cluster effects and are limited to regression and binary targets with only one clustering feature. We present MC-GMENN, a novel approach employing Monte Carlo techniques to train Generalized Mixed Effects Neural Networks. We empirically demonstrate that MC-GMENN outperforms existing mixed effects deep learning models in terms of generalization performance, time complexity, and quantification of inter-cluster variance. Additionally, MC-GMENN is applicable to a wide range of datasets, including multi-class classification tasks with multiple high-cardinality categorical features. For these datasets, we show that MC-GMENN outperforms conventional encoding and embedding methods, simultaneously offering a principled methodology for interpreting the effects of clustering patterns.
Empathy allows us to assume others are like us and have goals analogous to our own. This can also at times be applied to multi-agent games - e.g. Agent 1's attraction to green balls is analogous to Agent 2's attraction to red balls. Drawing inspiration from empathy, we propose EMOTE, a simple and explainable inverse reinforcement learning (IRL) approach designed to model another agent's action-value function and from it, infer a unique reward function. This is done by referencing the learning agent's own action value function, removing the need to maintain independent action-value estimates for the modelled agents whilst simultaneously addressing the ill-posed nature of IRL by inferring a unique reward function. We experiment on minigrid environments showing EMOTE: (a) produces more consistent reward estimates relative to other IRL baselines (b) is robust in scenarios with composite reward and action-value functions (c) produces human-interpretable states, helping to explain how the agent views other agents.
Compositionality is thought to be a key component of language, and various compositional benchmarks have been developed to empirically probe the compositional generalization of existing sequence processing models. These benchmarks often highlight failures of existing models, but it is not clear why these models fail in this way. In this paper, we seek to theoretically understand the role the compositional structure of the models plays in these failures and how this structure relates to their expressivity and sample complexity. We propose a general neuro-symbolic definition of compositional functions and their compositional complexity. We then show how various existing general and special purpose sequence processing models (such as recurrent, convolution and attention-based ones) fit this definition and use it to analyze their compositional complexity. Finally, we provide theoretical guarantees for the expressivity and systematic generalization of compositional models that explicitly depend on our proposed definition and highlighting factors which drive poor empirical performance.
Efficiency Calibration of Implicit Regularization in Deep Networks via Self-paced Curriculum-Driven Singular Value Selection
PDF ↗The generalization of neural networks has been a major focus of research in deep learning. It is often interpreted as an implicit bias towards solutions with specific properties. Especially, in practical applications, it has been observed that linear neural networks (LNN) tend to favor low-rank solutions for matrix completion tasks. However, most existing methods rely on increasing the depth of the neural network to enhance the low rank of solutions, resulting in higher complexity. In this paper, we propose a new explicit regularization method that calibrates the implicit bias towards low-rank trends in matrix completion tasks. Our approach automatically incorporates smaller singular values into the training process using a self-paced learning strategy, gradually restoring matrix information. By jointly using both implicit and explicit regularization, we effectively capture the low-rank structure of LNN and accelerate its convergence. We also analyze how our proposed penalty term interacts with implicit regularization and provide theoretical guarantees for our new model. To evaluate the effectiveness of our method, we conduct a series of experiments on both simulated and real-world data. Our experimental results clearly demonstrate that our method has better robustness and generalization ability compared with other methods.
Neural trees benefit from the high-level representation of neural networks and the interpretability of decision trees. Therefore, the existing works on neural trees perform outstandingly on various tasks such as architecture search. However, these works require every router to provide only one successor for each sample, causing the predictions to be dominated by the elite branch and its derivative architectures. To break this branch dominance, we propose the combinatorial routing neural tree method, termed CombRo. Unlike the previous methods employing unicast routing, CombRo performs multicast schema in each iteration, allowing the features to be routed to any combination of successors at every non-leaf. The weights of each architecture are then evaluated accordingly. We update the weights by training the routing subnetwork, and the architecture with the top weight is selected in the final step. We compare CombRo with the existing algorithms on 3 public image datasets, demonstrating its superior performance in terms of accuracy. Visualization results further validate the effectiveness of the multicast routing schema. Code is available at https://github.com/JiahaoLi-gdut/CombRo.
Deep Neural Networks (DNNs) can be represented as graphs whose links and vertices iteratively process data and solve tasks sub-optimally. Complex Network Theory (CNT), merging statistical physics with graph theory, provides a method for interpreting neural networks by analysing their weights and neuron structures. However, classic works adapt CNT metrics that only permit a topological analysis as they do not account for the effect of the input data. In addition, CNT metrics have been applied to a limited range of architectures, mainly including Fully Connected neural networks. In this work, we extend the existing CNT metrics with measures that sample from the DNNs' training distribution, shifting from a purely topological analysis to one that connects with the interpretability of deep learning. For the novel metrics, in addition to the existing ones, we provide a mathematical formalisation for Fully Connected, AutoEncoder, Convolutional and Recurrent neural networks, of which we vary the activation functions and the number of hidden layers. We show that these metrics differentiate DNNs based on the architecture, the number of hidden layers, and the activation function. Our contribution provides a method rooted in physics for interpreting DNNs that offers insights beyond the traditional input-output relationship and the CNT topological analysis.
In a regression task, a function is learned from labeled data to predict the labels at new data points. The goal is to achieve small prediction errors. In symbolic regression, the goal is more ambitious, namely, to learn an interpretable function that makes small prediction errors. This additional goal largely rules out the standard approach used in regression, that is, reducing the learning problem to learning parameters of an expansion of basis functions by optimization. Instead, symbolic regression methods search for a good solution in a space of symbolic expressions. To cope with the typically vast search space, most symbolic regression methods make implicit, or sometimes even explicit, assumptions about its structure. Here, we argue that the only obvious structure of the search space is that it contains small expressions, that is, expressions that can be decomposed into a few subexpressions. We show that systematically searching spaces of small expressions finds solutions that are more accurate and more robust against noise than those obtained by state-of-the-art symbolic regression methods. In particular, systematic search outperforms state-of-the-art symbolic regressors in terms of its ability to recover the true underlying symbolic expressions on established benchmark data sets.
Much research in machine learning involves finding appropriate inductive biases (e.g. convolutional neural networks, momentum-based optimizers, transformers) to promote generalization on tasks. However, quantification of the amount of inductive bias associated with these architectures and hyperparameters has been limited. We propose a novel method for efficiently computing the inductive bias required for generalization on a task with a fixed training data budget; formally, this corresponds to the amount of information required to specify well-generalizing models within a specific hypothesis space of models. Our approach involves modeling the loss distribution of random hypotheses drawn from a hypothesis space to estimate the required inductive bias for a task relative to these hypotheses. Unlike prior work, our method provides a direct estimate of inductive bias without using bounds and is applicable to diverse hypothesis spaces. Moreover, we derive approximation error bounds for our estimation approach in terms of the number of sampled hypotheses. Consistent with prior results, our empirical results demonstrate that higher dimensional tasks require greater inductive bias. We show that relative to other expressive model classes, neural networks as a model class encode large amounts of inductive bias. Furthermore, our measure quantifies the relative difference in inductive bias between different neural network architectures. Our proposed inductive bias metric provides an information-theoretic interpretation of the benefits of specific model architectures for certain tasks and provides a quantitative guide to developing tasks requiring greater inductive bias, thereby encouraging the development of more powerful inductive biases.
Interpretable Network Visualizations: A Human-in-the-Loop Approach for Post-hoc Explainability of CNN-based Image Classification
PDF ↗Transparency and explainability in image classification are essential for establishing trust in machine learning models and detecting biases and errors. State-of-the-art explainability methods generate saliency maps to show where a specific class is identified, without providing a detailed explanation of the model's decision process. Striving to address such a need, we introduce a post-hoc method that explains the entire feature extraction process of a Convolutional Neural Network. These explanations include a layer-wise representation of the features the model extracts from the input. Such features are represented as saliency maps generated by clustering and merging similar feature maps, to which we associate a weight derived by generalizing Grad-CAM for the proposed methodology. To further enhance these explanations, we include a set of textual labels collected through a gamified crowdsourcing activity and processed using NLP techniques and Sentence-BERT. Finally, we show an approach to generate global explanations by aggregating labels across multiple images.
Cutting the Black Box: Conceptual Interpretation of a Deep Neural Net with Multi-Modal Embeddings and Multi-Criteria Decision Aid
PDF ↗This paper tackles the concept-based explanation of neural models in computer vision, building upon the state of the art in Multi-Criteria Decision Aid (MCDA). The novelty of the approach is to leverage multi-modal embeddings from CLIP to bridge the gap between pixel-based and concept-based representations. The proposed Cut the Black Box (CB2) approach disentangles the latent representation of a trained pixel-based neural net, referred to as teacher model, along a 3-step process. Firstly, the pixel-based representation of the samples is mapped onto a conceptual representation using multi-modal embeddings. Secondly, an interpretable-by-design MCDA student model is trained by distillation from the teacher model, using the conceptual sample representation. Thirdly, the alignment of the teacher and student latent representations spells out the concepts relevant to explaining the teacher model. The empirical validation of the approach on ResNet, VGG, and VisionTransformer on Cifar-10, Cifar-100, Tiny ImageNet, and Fashion-MNIST showcases the effectiveness of the interpretations provided for the teacher models. The analysis reveals that decision-making predominantly relies on few concepts, thereby exposing potential bias in the teacher's decisions.