Linguistic commentary on LLMs, heavily influenced by the theoretical frameworks of de Saussure and Chomsky, is often speculative and unproductive. Critics challenge whether LLMs can legitimately model language, citing the need for "deep structure" or "grounding" to achieve an idealized linguistic "competence." We argue for a radical shift in perspective towards the empiricist principles of Witold Mańczak, a prominent general and historical linguist. He defines language not as a "system of signs" or a "computational system of the brain" but as the totality of all that is said and written. Above all, he identifies frequency of use of particular language elements as language's primary governing principle. Using his framework, we challenge prior critiques of LLMs and provide a constructive guide for designing, evaluating, and interpreting language models.

Large language models (LLMs) have shown remarkable capability in natural language tasks, yet debate persists on whether they truly comprehend deep structure (i.e., core semantics) or merely rely on surface structure (e.g., presentation format). Prior studies observe that LLMs' performance declines when intervening on surface structure, arguing their success relies on surface structure recognition. However, surface structure sensitivity does not prevent deep structure comprehension. Rigorously evaluating LLMs' capability requires analyzing both, yet deep structure is often overlooked. To this end, we assess LLMs' comprehension ability using causal mediation analysis, aiming to fully discover the capability of using both deep and surface structures. Specifically, we formulate the comprehension of deep structure as direct causal effect (DCE) and that of surface structure as indirect causal effect (ICE), respectively. To address the non-estimability of original DCE and ICE -- stemming from the infeasibility of isolating mutual influences of deep and surface structures, we develop the corresponding quantifiable surrogates, including approximated DCE (ADCE) and approximated ICE (AICE). We further apply the ADCE to evaluate a series of mainstream LLMs, showing that most of them exhibit deep structure comprehension ability, which grows along with the prediction accuracy. Comparing ADCE and AICE demonstrates closed-source LLMs rely more on deep structure, while open-source LLMs are more surface-sensitive, which decreases with model scale. Theoretically, ADCE is a bidirectional evaluation, which measures both the sufficiency and necessity of deep structure changes in causing output variations, thus offering a more comprehensive assessment than accuracy, a common evaluation in LLMs. Our work provides new insights into LLMs' deep structure comprehension and offers novel methods for LLMs evaluation.

Clustering is a fundamental unsupervised representation learning task with wide application in computer vision and pattern recognition. Deep clustering utilizes deep neural networks to learn latent representation, which is suitable for clustering. However, previous deep clustering methods, especially image clustering, focus on the features of the data itself and ignore the relationship between the data, which is crucial for clustering. In this paper, we propose a novel Deep Structure and Attention aware Subspace Clustering (DSASC), which simultaneously considers data content and structure information. We use a vision transformer to extract features, and the extracted features are divided into two parts, structure features, and content features. The two features are used to learn a more efficient subspace structure for spectral clustering. Extensive experimental results demonstrate that our method significantly outperforms state-of-the-art methods. Our code will be available at https://github.com/cs-whh/DSASC

Clustering has been one of the most critical unsupervised learning techniques that has been widely applied in data mining problems. As one of its branches, graph clustering enjoys its popularity due to its appealing performance and strong theoretical supports. However, the eigen-decomposition problems involved are computationally expensive. In this paper, we propose a deep structure with a linear coder as the building block for fast graph clustering, called Deep Linear Coding (DLC). Different from conventional coding schemes, we jointly learn the feature transform function and discriminative codings, and guarantee that the learned codes are robust in spite of local distortions. In addition, we use the proposed linear coders as the building blocks to formulate a deep structure to further refine features in a layerwise fashion. Extensive experiments on clustering tasks demonstrate that our method performs well in terms of both time complexity and clustering accuracy. On a large-scale benchmark dataset (580K), our method runs 1500 times faster than the original spectral clustering.

Deep Optimisation (DO) combines evolutionary search with Deep Neural Networks (DNNs) in a novel way - not for optimising a learning algorithm, but for finding a solution to an optimisation problem. Deep learning has been successfully applied to classification, regression, decision and generative tasks and in this paper we extend its application to solving optimisation problems. Model Building Optimisation Algorithms (MBOAs), a branch of evolutionary algorithms, have been successful in combining machine learning methods and evolutionary search but, until now, they have not utilised DNNs. DO is the first algorithm to use a DNN to learn and exploit the problem structure to adapt the variation operator (changing the neighbourhood structure of the search process). We demonstrate the performance of DO using two theoretical optimisation problems within the MAXSAT class. The Hierarchical Transformation Optimisation Problem (HTOP) has controllable deep structure that provides a clear evaluation of how DO works and why using a layerwise technique is essential for learning and exploiting problem structure. The Parity Modular Constraint Problem (MCparity) is a simplistic example of a problem containing higher-order dependencies (greater than pairwise) which DO can solve and state of the art MBOAs cannot. Further, we show that DO can exploit deep structure in TSP instances. Together these results show that there exists problems that DO can find and exploit deep problem structure that other algorithms cannot. Making this connection between DNNs and optimisation allows for the utilisation of advanced tools applicable to DNNs that current MBOAs are unable to use.

Multi-View Clustering (MVC) has garnered more attention recently since many real-world data are comprised of different representations or views. The key is to explore complementary information to benefit the clustering problem. In this paper, we present a deep matrix factorization framework for MVC, where semi-nonnegative matrix factorization is adopted to learn the hierarchical semantics of multi-view data in a layer-wise fashion. To maximize the mutual information from each view, we enforce the non-negative representation of each view in the final layer to be the same. Furthermore, to respect the intrinsic geometric structure in each view data, graph regularizers are introduced to couple the output representation of deep structures. As a non-trivial contribution, we provide the solution based on alternating minimization strategy, followed by a theoretical proof of convergence. The superior experimental results on three face benchmarks show the effectiveness of the proposed deep matrix factorization model.

Clustering has been one of the most critical unsupervised learning techniques that has been widely applied in data mining problems. As one of its branches, graph clustering enjoys its popularity due to its appealing performance and strong theoretical supports. However, the eigen-decomposition problems involved are computationally expensive. In this paper, we propose a deep structure with a linear coder as the building block for fast graph clustering, called Deep Linear Coding (DLC). Different from conventional coding schemes, we jointly learn the feature transform function and discriminative codings, and guarantee that the learned codes are robust in spite of local distortions. In addition, we use the proposed linear coders as the building blocks to formulate a deep structure to further refine features in a layerwise fashion. Extensive experiments on clustering tasks demonstrate that our method performs well in terms of both time complexity and clustering accuracy. On a large-scale benchmark dataset (580K), our method runs 1500 times faster than the original spectral clustering.

Recent researches on transfer learning exploit deep structures for discriminative feature representation to tackle cross-domain disparity. However, few of them are able to joint feature learning and knowledge transfer in a unified deep framework. In this paper, we develop a novel approach, called Deep Low-Rank Coding (DLRC), for transfer learning. Specifically, discriminative low-rank coding is achieved in the guidance of an iterative supervised structure term for each single layer. In this way, both marginal and conditional distributions between two domains intend to be mitigated. In addition, a marginalized denoising feature transformation is employed to guarantee the learned single-layer low-rank coding to be robust despite of corruptions or noises. Finally, by stacking multiple layers of low-rank codings, we manage to learn robust cross-domain features from coarse to fine. Experimental results on several benchmarks have demonstrated the effectiveness of our proposed algorithm on facilitating the recognition performance for the target domain.

Multimodal learning with deep Boltzmann machines (DBMs) is an generative approach to fuse multimodal inputs, and can learn the shared representation via Contrastive Divergence (CD) for classification and information retrieval tasks. However, it is a 2-fan DBM model, and cannot effectively handle multiple prediction tasks. Moreover, this model cannot recover the hidden representations well by sampling from the conditional distribution when more than one modalities are missing. In this paper, we propose a K-fan deep structure model, which can handle the multi-input and muti-output learning problems effectively. In particular, the deep structure has K-branch for different inputs where each branch can be composed of a multi-layer deep model, and a shared representation is learned in an discriminative manner to tackle multimodal tasks. Given the deep structure, we propose two objective functions to handle two multi-input and multi-output tasks: joint visual restoration and labeling, and the multi-view multi-calss object recognition tasks. To estimate the model parameters, we initialize the deep model parameters with CD to maximize the joint distribution, and then we use backpropagation to update the model according to specific objective function. The experimental results demonstrate that the model can effectively leverages multi-source information and predict multiple tasks well over competitive baselines.

Recently deep learning has been successfully adopted in many applications such as speech recognition and image classification. In this work, we explore the possibility of employing deep learning in graph clustering. We propose a simple method, which first learns a nonlinear embedding of the original graph by stacked autoencoder, and then runs $k$-means algorithm on the embedding to obtain the clustering result. We show that this simple method has solid theoretical foundation, due to the similarity between autoencoder and spectral clustering in terms of what they actually optimize. Then, we demonstrate that the proposed method is more efficient and flexible than spectral clustering. First, the computational complexity of autoencoder is much lower than spectral clustering: the former can be linear to the number of nodes in a sparse graph while the latter is super quadratic due to eigenvalue decomposition. Second, when additional sparsity constraint is imposed, we can simply employ the sparse autoencoder developed in the literature of deep learning; however, it is non-straightforward to implement a sparse spectral method. The experimental results on various graph datasets show that the proposed method significantly outperforms conventional spectral clustering which clearly indicates the effectiveness of deep learning in graph clustering.