Transformer-based models have dominated natural language processing and other areas in the last few years due to their superior (zero-shot) performance on benchmark datasets. However, these models are poorly understood due to their complexity and size. While probing-based methods are widely used to understand specific properties, the structures of the representation space are not systematically characterized; consequently, it is unclear how such models generalize and overgeneralize to new inputs beyond datasets. In this paper, based on a new gradient descent optimization method, we are able to explore the embedding space of a commonly used vision-language model. Using the Imagenette dataset, we show that while the model achieves over 99\% zero-shot classification performance, it fails systematic evaluations completely. Using a linear approximation, we provide a framework to explain the striking differences. We have also obtained similar results using a different model to support that our results are applicable to other transformer models with continuous inputs. We also propose a robust way to detect the modified images.
Pre-trained large foundation models play a central role in the recent surge of artificial intelligence, resulting in fine-tuned models with remarkable abilities when measured on benchmark datasets, standard exams, and applications. Due to their inherent complexity, these models are not well understood. While small adversarial inputs to such models are well known, the structures of the representation space are not well characterized despite their fundamental importance. In this paper, using the vision transformers as an example due to the continuous nature of their input space, we show via analyses and systematic experiments that the representation space consists of large piecewise linear subspaces where there exist very different inputs sharing the same representations, and at the same time, local normal spaces where there are visually indistinguishable inputs having very different representations. The empirical results are further verified using the local directional estimations of the Lipschitz constants of the underlying models. Consequently, the resulting representations change the results of downstream models, and such models are subject to overgeneralization and with limited semantically meaningful generalization capability.
Link prediction is a crucial research area in knowledge graphs, with many downstream applications. In many real-world scenarios, inductive link prediction is required, where predictions have to be made among unseen entities. Embedding-based models usually need fine-tuning on new entity embeddings, and hence are difficult to be directly applied to inductive link prediction tasks. Logical rules captured by rule-based models can be directly applied to new entities with the same graph typologies, but the captured rules are discrete and usually lack generosity. Graph neural networks (GNNs) can generalize topological information to new graphs taking advantage of deep neural networks, which however may still need fine-tuning on new entity embeddings. In this paper, we propose SiaILP, a path-based model for inductive link prediction using siamese neural networks. Our model only depends on relation and path embeddings, which can be generalized to new entities without fine-tuning. Experiments show that our model achieves several new state-of-the-art performances in link prediction tasks using inductive versions of WN18RR, FB15k-237, and Nell995.
Graph-based semi-supervised learning (GSSL) has been used successfully in various applications. Existing methods leverage the graph structure and labeled samples for classification. Label Propagation (LP) and Graph Neural Networks (GNNs) both iteratively pass messages on graphs, where LP propagates node labels through edges and GNN aggregates node features from the neighborhood. Recently, combining LP and GNN has led to improved performance. However, utilizing labels and features jointly in higher-order graphs has not been explored. Therefore, we propose Nonlinear Correct and Smooth (NLCS), which improves the existing post-processing approach by incorporating non-linearity and higher-order representation into the residual propagation to handle intricate node relationships effectively. Systematic evaluations show that our method achieves remarkable average improvements of 13.71% over base prediction and 2.16% over the state-of-the-art post-processing method on six commonly used datasets. Comparisons and analyses show our method effectively utilizes labels and features jointly in higher-order graphs to resolve challenging graph relationships.
Stock price movement prediction is a challenging and essential problem in finance. While it is well established in modern behavioral finance that the share prices of related stocks often move after the release of news via reactions and overreactions of investors, how to capture the relationships between price movements and news articles via quantitative models is an active area research; existing models have achieved success with variable degrees. In this paper, we propose to improve stock price movement classification using news articles by incorporating regularization and optimization techniques from deep learning. More specifically, we capture the dependencies between news articles and stocks through embeddings and bidirectional recurrent neural networks as in recent models. We further incorporate weight decay, batch normalization, dropout, and label smoothing to improve the generalization of the trained models. To handle high fluctuations of validation accuracy of batch normalization, we propose dual-phase training to realize the improvements reliably. Our experimental results on a commonly used dataset show significant improvements, achieving average accuracy of 80.7% on the test set, which is more than 10.0% absolute improvement over existing models. Our ablation studies show batch normalization and label smoothing are most effective, leading to 6.0% and 3.4% absolute improvement, respectively on average.
Recently, multi-layer perceptrons (MLPs) with ReLU activations have enabled new photo-realistic rendering techniques by encoding scene properties using their weights. For these models, termed coordinate based MLPs, sinusoidal encodings are necessary in allowing for convergence to the high frequency components of the signal due to their severe spectral bias. Previous work has explained this phenomenon using Neural Tangent Kernel (NTK) and Fourier analysis. However, the kernel regime does not expose the properties of the network that induce this behavior, and the Fourier decomposition is global, not allowing for insight on the network's local dynamics. A new interpretation of spectral bias directly through ReLU network computations would expose their limitations in dense settings, while providing a clearer explanation as to how this behavior emerges during the learning process. In this paper, we provide the first study of spectral bias in a coordinate based MLP through its activation regions and gradient descent dynamics, specifically using gradient confusion. We relate the confusion between inputs to the distinctiveness of their activation patterns, and find higher amounts of confusion when expressive power is limited. This leads to slower convergence to the high frequency components of the signal, which is magnified by the density of coordinates. Additionally, this method allows us to analyze the properties of the activation regions as spectral bias is reduced, in which we find distinct dynamics.
In this paper, we provide a novel way to generate low-dimension (dense) vector embeddings for the noun and verb synsets in WordNet, so that the hypernym-hyponym tree structure is preserved in the embeddings. We call this embedding the sense spectrum (and sense spectra for embeddings). In order to create suitable labels for the training of sense spectra, we designed a new similarity measurement for noun and verb synsets in WordNet. We call this similarity measurement the hypernym intersection similarity (HIS), since it compares the common and unique hypernyms between two synsets. Our experiments show that on the noun and verb pairs of the SimLex-999 dataset, HIS outperforms the three similarity measurements in WordNet. Moreover, to the best of our knowledge, the sense spectra is the first dense embedding system that can explicitly and completely measure the hypernym-hyponym relationship in WordNet.
To improve the generalization of the representations for natural language processing tasks, words are commonly represented using vectors, where distances among the vectors are related to the similarity of the words. While word2vec, the state-of-the-art implementation of the skip-gram model, is widely used and improves the performance of many natural language processing tasks, its mechanism is not yet well understood. In this work, we derive the learning rules for the skip-gram model and establish their close relationship to competitive learning. In addition, we provide the global optimal solution constraints for the skip-gram model and validate them by experimental results.
Understanding the underlying mechanisms that enable the empirical successes of deep neural networks is essential for further improving their performance and explaining such networks. Towards this goal, a specific question is how to explain the "surprising" behavior of the same over-parametrized deep neural networks that can generalize well on real datasets and at the same time "memorize" training samples when the labels are randomized. In this paper, we demonstrate that deep ReLU networks generalize from training samples to new points via piece-wise linear interpolation. We provide a quantified analysis on the generalization ability of a deep ReLU network: Given a fixed point $\mathbf{x}$ and a fixed direction in the input space $\mathcal{S}$, there is always a segment such that any point on the segment will be classified the same as the fixed point $\mathbf{x}$. We call this segment the $generalization \ interval$. We show that the generalization intervals of a ReLU network behave similarly along pairwise directions between samples of the same label in both real and random cases on the MNIST and CIFAR-10 datasets. This result suggests that the same interpolation mechanism is used in both cases. Additionally, for datasets using real labels, such networks provide a good approximation of the underlying manifold in the data, where the changes are much smaller along tangent directions than along normal directions. On the other hand, however, for datasets with random labels, generalization intervals along mid-lines of triangles with the same label are much smaller than those on the datasets with real labels, suggesting different behaviors along other directions. Our systematic experiments demonstrate for the first time that such deep neural networks generalize through the same interpolation and explain the differences between their performance on datasets with real and random labels.
Deep neural networks have achieved remarkable success in challenging tasks. However, the black-box approach of training and testing of such networks is not acceptable to critical applications. In particular, the existence of adversarial examples and their overgeneralization to irrelevant inputs makes it difficult, if not impossible, to explain decisions by commonly used neural networks. In this paper, we analyze the underlying mechanism of generalization of deep neural networks and propose an ($n$, $k$) consensus algorithm to be insensitive to adversarial examples and at the same time be able to reject irrelevant samples. Furthermore, the consensus algorithm is able to improve classification accuracy by using multiple trained deep neural networks. To handle the complexity of deep neural networks, we cluster linear approximations and use cluster means to capture feature importance. Due to weight symmetry, a small number of clusters are sufficient to produce a robust interpretation. Experimental results on a health dataset show the effectiveness of our algorithm in enhancing the prediction accuracy and interpretability of deep neural network models on one-year patient mortality prediction.