The existence of external (``side'') semantic knowledge has been shown to result in more expressive computational event models. To enable the use of side information that may be noisy or missing, we propose a semi-supervised information bottleneck-based discrete latent variable model. We reparameterize the model's discrete variables with auxiliary continuous latent variables and a light-weight hierarchical structure. Our model is learned to minimize the mutual information between the observed data and optional side knowledge that is not already captured by the new, auxiliary variables. We theoretically show that our approach generalizes past approaches, and perform an empirical case study of our approach on event modeling. We corroborate our theoretical results with strong empirical experiments, showing that the proposed method outperforms previous proposed approaches on multiple datasets.

Generating text with autoregressive language models (LMs) is of great importance to many natural language processing (NLP) applications. Previous solutions for this task often produce text that contains degenerative expressions or lacks semantic consistency. Recently, Su et al. introduced a new decoding method, contrastive search, based on the isotropic representation space of the language model and obtained new state of the art on various benchmarks. Additionally, Su et al. argued that the representations of autoregressive LMs (e.g. GPT-2) are intrinsically anisotropic which is also shared by previous studies. Therefore, to ensure the language model follows an isotropic distribution, Su et al. proposed a contrastive learning scheme, SimCTG, which calibrates the language model's representations through additional training. In this study, we first answer the question: "Are autoregressive LMs really anisotropic?". To this end, we extensively evaluate the isotropy of LMs across 16 major languages. Surprisingly, we find that the anisotropic problem only exists in the two specific English GPT-2-small/medium models. On the other hand, all other evaluated LMs are naturally isotropic which is in contrast to the conclusion drawn by previous studies. Based on our findings, we further assess the contrastive search decoding method using off-the-shelf LMs on four generation tasks across 16 languages. Our experimental results demonstrate that contrastive search significantly outperforms previous decoding methods without any additional training. More notably, on 12 out of the 16 evaluated languages, contrastive search performs comparably with human-level performances as judged by human evaluations. Our code and other related resources are publicly available at https://github.com/yxuansu/Contrastive_Search_Is_What_You_Need.

Activation functions can have a significant impact on reducing the topological complexity of input data and therefore improve the performance of the model. Selecting a suitable activation function is an essential step in neural model design. However, the choice of activation function is seldom discussed or explored in Transformer-based language models. Their activation functions are chosen beforehand and then remain fixed from pre-training to fine-tuning. As a result, the inductive biases they imposed on models cannot be adjusted during this long life cycle. Moreover, subsequently developed models (e.g., RoBERTa, BART, and GPT-3) often follow up prior work (e.g., BERT) to use the same activation function without justification. In this paper, we investigate the effectiveness of using Rational Activation Function (RAF), a learnable activation function, in the Transformer architecture. In contrast to conventional, predefined activation functions, RAFs can adaptively learn optimal activation functions during training according to input data. Our experiments show the RAF-based Transformer (RAFT) achieves a lower validation perplexity than a vanilla BERT with the GELU function. We further evaluate RAFT on downstream tasks in low- and full-data settings. Our results show that RAFT outperforms the counterpart model across the majority of tasks and settings. For instance, RAFT outperforms vanilla BERT on the GLUE benchmark by 5.71 points on average in low-data scenario (where 100 training examples are available) and by 2.05 points on SQuAD in full-data setting. Analysis of the shapes of learned RAFs further unveils that they substantially vary between different layers of the pre-trained model and mostly look very different from conventional activation functions. RAFT opens a new research direction for analyzing and interpreting pre-trained models according to the learned activation functions.

Originally published on Towards AI. With ChatGPT blowing the internet, we are at a critical juncture that demands us to again ask hard questions about the impact of AI models on society, a conversation that starts but never ends. In this article, I aim to bring attention to the importance of knowing that, even though large AI models are impressive, there are often unacknowledged costs behind them. It is like saying " data is the new oil" to describe its value, but this analogy often ignores the costs of the oil and mining industries. To understand what AI is made from, we need to leave Silicon Valley and go to the place where the stuff for the AI industry is made. The term "artificial intelligence" may evoke the ideas of algorithms and data, but it is powered by the rare earth's minerals and resources that make up the computing components [1].

The study of stability and sensitivity of statistical methods or algorithms with respect to their data is an important problem in machine learning and statistics. The performance of the algorithm under resampling of the data is a fundamental way to measure its stability and is closely related to generalization or privacy of the algorithm. In this paper, we study the resampling sensitivity for the principal component analysis (PCA). Given an $ n \times p $ random matrix $ \mathbf{X} $, let $ \mathbf{X}^{[k]} $ be the matrix obtained from $ \mathbf{X} $ by resampling $ k $ randomly chosen entries of $ \mathbf{X} $. Let $ \mathbf{v} $ and $ \mathbf{v}^{[k]} $ denote the principal components of $ \mathbf{X} $ and $ \mathbf{X}^{[k]} $. In the proportional growth regime $ p/n \to \xi \in (0,1] $, we establish the sharp threshold for the sensitivity/stability transition of PCA. When $ k \gg n^{5/3} $, the principal components $ \mathbf{v} $ and $ \mathbf{v}^{[k]} $ are asymptotically orthogonal. On the other hand, when $ k \ll n^{5/3} $, the principal components $ \mathbf{v} $ and $ \mathbf{v}^{[k]} $ are asymptotically colinear. In words, we show that PCA is sensitive to the input data in the sense that resampling even a negligible portion of the input may completely change the output.

Understanding agent behaviour in Multi-Agent Systems (MAS) is an important problem in domains such as autonomous driving, disaster response, and sports analytics. Existing MAS problems typically use uniform timesteps with observations for all agents. In this work, we analyse the problem of agent location imputation, specifically posed in environments with non-uniform timesteps and limited agent observability (~95% missing values). Our approach uses Long Short-Term Memory and Graph Neural Network components to learn temporal and inter-agent patterns to predict the location of all agents at every timestep. We apply this to the domain of football (soccer) by imputing the location of all players in a game from sparse event data (e.g., shots and passes). Our model estimates player locations to within ~6.9m; a ~62% reduction in error from the best performing baseline. This approach facilitates downstream analysis tasks such as player physical metrics, player coverage, and team pitch control. Existing solutions to these tasks often require optical tracking data, which is expensive to obtain and only available to elite clubs. By imputing player locations from easy to obtain event data, we increase the accessibility of downstream tasks.

Predicting missing links between entities in a knowledge graph is a fundamental task to deal with the incompleteness of data on the Web. Knowledge graph embeddings map nodes into a vector space to predict new links, scoring them according to geometric criteria. Relations in the graph may follow patterns that can be learned, e.g., some relations might be symmetric and others might be hierarchical. However, the learning capability of different embedding models varies for each pattern and, so far, no single model can learn all patterns equally well. In this paper, we combine the query representations from several models in a unified one to incorporate patterns that are independently captured by each model. Our combination uses attention to select the most suitable model to answer each query. The models are also mapped onto a non-Euclidean manifold, the Poincar\'e ball, to capture structural patterns, such as hierarchies, besides relational patterns, such as symmetry. We prove that our combination provides a higher expressiveness and inference power than each model on its own. As a result, the combined model can learn relational and structural patterns. We conduct extensive experimental analysis with various link prediction benchmarks showing that the combined model outperforms individual models, including state-of-the-art approaches.

Existing language and vision models achieve impressive performance in image-text understanding. Yet, it is an open question to what extent they can be used for language understanding in 3D environments and whether they implicitly acquire 3D object knowledge, e.g. about different views of an object. In this paper, we investigate whether a state-of-the-art language and vision model, CLIP, is able to ground perspective descriptions of a 3D object and identify canonical views of common objects based on text queries. We present an evaluation framework that uses a circling camera around a 3D object to generate images from different viewpoints and evaluate them in terms of their similarity to natural language descriptions. We find that a pre-trained CLIP model performs poorly on most canonical views and that fine-tuning using hard negative sampling and random contrasting yields good results even under conditions with little available training data.

In this paper, we study a spiked Wigner problem with an inhomogeneous noise profile. Our aim in this problem is to recover the signal passed through an inhomogeneous low-rank matrix channel. While the information-theoretic performances are well-known, we focus on the algorithmic problem. We derive an approximate message-passing algorithm (AMP) for the inhomogeneous problem and show that its rigorous state evolution coincides with the information-theoretic optimal Bayes fixed-point equations. We identify in particular the existence of a statistical-to-computational gap where known algorithms require a signal-to-noise ratio bigger than the information-theoretic threshold to perform better than random. Finally, from the adapted AMP iteration we deduce a simple and efficient spectral method that can be used to recover the transition for matrices with general variance profiles. This spectral method matches the conjectured optimal computational phase transition.

Generative Artificial Intelligence (AI) has enabled the development of sophisticated models that are capable of producing high-caliber text, images, and other outputs through the utilization of large pre-trained models. Nevertheless, assessing the quality of the generation is an even more arduous task than the generation itself, and this issue has not been given adequate consideration recently. This paper proposes a novel evaluation framework, GPTScore, which utilizes the emergent abilities (e.g., zero-shot instruction) of generative pre-trained models to score generated texts. There are 19 pre-trained models explored in this paper, ranging in size from 80M (e.g., FLAN-T5-small) to 175B (e.g., GPT3). Experimental results on four text generation tasks, 22 evaluation aspects, and corresponding 37 datasets demonstrate that this approach can effectively allow us to achieve what one desires to evaluate for texts simply by natural language instructions. This nature helps us overcome several long-standing challenges in text evaluation--how to achieve customized, multi-faceted evaluation without the need for annotated samples. We make our code publicly available at https://github.com/jinlanfu/GPTScore.