Stochastic variational Bayes algorithms have become very popular in the machine learning literature, particularly in the context of nonparametric Bayesian inference. These algorithms replace the true but intractable posterior distribution with the best (in the sense of Kullback-Leibler divergence) member of a tractable family of distributions, using stochastic gradient algorithms to perform the optimization step. stochastic variational Bayes inference implicitly trades off computational speed for accuracy, but the loss of accuracy is highly model (and even dataset) specific. In this paper we carry out an empirical evaluation of this trade off in the context of stochastic blockmodels, which are a widely used class of probabilistic models for network and relational data. Our experiments indicate that, in the context of stochastic blockmodels, relatively large subsamples are required for these algorithms to find accurate approximations of the posterior, and that even then the quality of the approximations provided by stochastic gradient variational algorithms can be highly variable.
相關內容
The Softmax attention mechanism in Transformer models is notoriously computationally expensive, particularly due to its quadratic complexity, posing significant challenges in vision applications. In contrast, linear attention provides a far more efficient solution by reducing the complexity to linear levels. However, compared to Softmax attention, linear attention often experiences significant performance degradation. Our experiments indicate that this performance drop is due to the low-rank nature of linear attention's feature map, which hinders its ability to adequately model complex spatial information. In this paper, to break the low-rank dilemma of linear attention, we conduct rank analysis from two perspectives: the KV buffer and the output features. Consequently, we introduce Rank-Augmented Linear Attention (RALA), which rivals the performance of Softmax attention while maintaining linear complexity and high efficiency. Based on RALA, we construct the Rank-Augmented Vision Linear Transformer (RAVLT). Extensive experiments demonstrate that RAVLT achieves excellent performance across various vision tasks. Specifically, without using any additional labels, data, or supervision during training, RAVLT achieves an 84.4% Top-1 accuracy on ImageNet-1k with only 26M parameters and 4.6G FLOPs. This result significantly surpasses previous linear attention mechanisms, fully illustrating the potential of RALA. Code will be available at //github.com/qhfan/RALA.
Large language models can absorb a massive amount of knowledge through pretraining, but pretraining is inefficient for acquiring long-tailed or specialized facts. Therefore, fine-tuning on specialized or new knowledge that reflects changes in the world has become popular, though it risks disrupting the model's original capabilities. We study this fragility in the context of continual memorization, where the model is trained on a small set of long-tail factoids (factual associations) and must retain these factoids after multiple stages of subsequent training on other datasets. Through extensive experiments, we show that LLMs suffer from forgetting across a wide range of subsequent tasks, and simple replay techniques do not fully prevent forgetting, especially when the factoid datasets are trained in the later stages. We posit that there are two ways to alleviate forgetting: 1) protect the memorization process as the model learns the factoids, or 2) reduce interference from training in later stages. With this insight, we develop an effective mitigation strategy: REMIX (Random and Generic Data Mixing). REMIX prevents forgetting by mixing generic data sampled from pretraining corpora or even randomly generated word sequences during each stage, despite being unrelated to the memorized factoids in the first stage. REMIX can recover performance from severe forgetting, often outperforming replay-based methods that have access to the factoids from the first stage. We then analyze how REMIX alters the learning process and find that successful forgetting prevention is associated with a pattern: the model stores factoids in earlier layers than usual and diversifies the set of layers that store these factoids. The efficacy of REMIX invites further investigation into the underlying dynamics of memorization and forgetting, opening exciting possibilities for future research.
We present a technique and benchmark dataset for estimating the relative 3D orientation between a pair of Internet images captured in an extreme setting, where the images have limited or non-overlapping field of views. Prior work targeting extreme rotation estimation assume constrained 3D environments and emulate perspective images by cropping regions from panoramic views. However, real images captured in the wild are highly diverse, exhibiting variation in both appearance and camera intrinsics. In this work, we propose a Transformer-based method for estimating relative rotations in extreme real-world settings, and contribute the ExtremeLandmarkPairs dataset, assembled from scene-level Internet photo collections. Our evaluation demonstrates that our approach succeeds in estimating the relative rotations in a wide variety of extremeview Internet image pairs, outperforming various baselines, including dedicated rotation estimation techniques and contemporary 3D reconstruction methods.
It has been shown that one can design distributed algorithms that are (nearly) singularly optimal, meaning they simultaneously achieve optimal time and message complexity (within polylogarithmic factors), for several fundamental global problems such as broadcast, leader election, and spanning tree construction, under the $\text{KT}_0$ assumption. With this assumption, nodes have initial knowledge only of themselves, not their neighbors. In this case the time and message lower bounds are $\Omega(D)$ and $\Omega(m)$, respectively, where $D$ is the diameter of the network and $m$ is the number of edges, and there exist (even) deterministic algorithms that simultaneously match these bounds. On the other hand, under the $\text{KT}_1$ assumption, whereby each node has initial knowledge of itself and the identifiers of its neighbors, the situation is not clear. For the $\text{KT}_1$ CONGEST model (where messages are of small size), King, Kutten, and Thorup (KKT) showed that one can solve several fundamental global problems (with the notable exception of BFS tree construction) such as broadcast, leader election, and spanning tree construction with $\tilde{O}(n)$ message complexity ($n$ is the network size), which can be significantly smaller than $m$. Randomization is crucial in obtaining this result. While the message complexity of the KKT result is near-optimal, its time complexity is $\tilde{O}(n)$ rounds, which is far from the standard lower bound of $\Omega(D)$. In this paper, we show that in the $\text{KT}_1$ LOCAL model (where message sizes are not restricted), singular optimality is achievable. Our main result is that all global problems, including BFS tree construction, can be solved in $\tilde{O}(D)$ rounds and $\tilde{O}(n)$ messages, where both bounds are optimal up to polylogarithmic factors. Moreover, we show that this can be achieved deterministically.
Recent papers in the graph machine learning literature have introduced a number of approaches for hyperbolic representation learning. The asserted benefits are improved performance on a variety of graph tasks, node classification and link prediction included. Claims have also been made about the geometric suitability of particular hierarchical graph datasets to representation in hyperbolic space. Despite these claims, our work makes a surprising discovery: when simple Euclidean models with comparable numbers of parameters are properly trained in the same environment, in most cases, they perform as well, if not better, than all introduced hyperbolic graph representation learning models, even on graph datasets previously claimed to be the most hyperbolic as measured by Gromov $\delta$-hyperbolicity (i.e., perfect trees). This observation gives rise to a simple question: how can this be? We answer this question by taking a careful look at the field of hyperbolic graph representation learning as it stands today, and find that a number of papers fail to diligently present baselines, make faulty modelling assumptions when constructing algorithms, and use misleading metrics to quantify geometry of graph datasets. We take a closer look at each of these three problems, elucidate the issues, perform an analysis of methods, and introduce a parametric family of benchmark datasets to ascertain the applicability of (hyperbolic) graph neural networks.
The decidability of the reachability problem for finitary PCF has been used as a theoretical basis for fully automated verification tools for functional programs. The reachability problem, however, often becomes undecidable for a slight extension of finitary PCF with side effects, such as exceptions, algebraic effects, and references, which hindered the extension of the above verification tools for supporting functional programs with side effects. In this paper, we first give simple proofs of the undecidability of four extensions of finitary PCF, which would help us understand and analyze the source of undecidability. We then focus on an extension with references, and give a decidable fragment using a type system. To our knowledge, this is the first non-trivial decidable fragment that features higher-order recursive functions containing reference cells.
Embeddings have become a cornerstone in the functionality of large language models (LLMs) due to their ability to transform text data into rich, dense numerical representations that capture semantic and syntactic properties. These embedding vector databases serve as the long-term memory of LLMs, enabling efficient handling of a wide range of natural language processing tasks. However, the surge in popularity of embedding vector databases in LLMs has been accompanied by significant concerns about privacy leakage. Embedding vector databases are particularly vulnerable to embedding inversion attacks, where adversaries can exploit the embeddings to reverse-engineer and extract sensitive information from the original text data. Existing defense mechanisms have shown limitations, often struggling to balance security with the performance of downstream tasks. To address these challenges, we introduce Eguard, a novel defense mechanism designed to mitigate embedding inversion attacks. Eguard employs a transformer-based projection network and text mutual information optimization to safeguard embeddings while preserving the utility of LLMs. Our approach significantly reduces privacy risks, protecting over 95% of tokens from inversion while maintaining high performance across downstream tasks consistent with original embeddings.
Pre-trained Language Models (PLMs) which are trained on large text corpus via self-supervised learning method, have yielded promising performance on various tasks in Natural Language Processing (NLP). However, though PLMs with huge parameters can effectively possess rich knowledge learned from massive training text and benefit downstream tasks at the fine-tuning stage, they still have some limitations such as poor reasoning ability due to the lack of external knowledge. Research has been dedicated to incorporating knowledge into PLMs to tackle these issues. In this paper, we present a comprehensive review of Knowledge-Enhanced Pre-trained Language Models (KE-PLMs) to provide a clear insight into this thriving field. We introduce appropriate taxonomies respectively for Natural Language Understanding (NLU) and Natural Language Generation (NLG) to highlight these two main tasks of NLP. For NLU, we divide the types of knowledge into four categories: linguistic knowledge, text knowledge, knowledge graph (KG), and rule knowledge. The KE-PLMs for NLG are categorized into KG-based and retrieval-based methods. Finally, we point out some promising future directions of KE-PLMs.
The generalization mystery in deep learning is the following: Why do over-parameterized neural networks trained with gradient descent (GD) generalize well on real datasets even though they are capable of fitting random datasets of comparable size? Furthermore, from among all solutions that fit the training data, how does GD find one that generalizes well (when such a well-generalizing solution exists)? We argue that the answer to both questions lies in the interaction of the gradients of different examples during training. Intuitively, if the per-example gradients are well-aligned, that is, if they are coherent, then one may expect GD to be (algorithmically) stable, and hence generalize well. We formalize this argument with an easy to compute and interpretable metric for coherence, and show that the metric takes on very different values on real and random datasets for several common vision networks. The theory also explains a number of other phenomena in deep learning, such as why some examples are reliably learned earlier than others, why early stopping works, and why it is possible to learn from noisy labels. Moreover, since the theory provides a causal explanation of how GD finds a well-generalizing solution when one exists, it motivates a class of simple modifications to GD that attenuate memorization and improve generalization. Generalization in deep learning is an extremely broad phenomenon, and therefore, it requires an equally general explanation. We conclude with a survey of alternative lines of attack on this problem, and argue that the proposed approach is the most viable one on this basis.
Recent contrastive representation learning methods rely on estimating mutual information (MI) between multiple views of an underlying context. E.g., we can derive multiple views of a given image by applying data augmentation, or we can split a sequence into views comprising the past and future of some step in the sequence. Contrastive lower bounds on MI are easy to optimize, but have a strong underestimation bias when estimating large amounts of MI. We propose decomposing the full MI estimation problem into a sum of smaller estimation problems by splitting one of the views into progressively more informed subviews and by applying the chain rule on MI between the decomposed views. This expression contains a sum of unconditional and conditional MI terms, each measuring modest chunks of the total MI, which facilitates approximation via contrastive bounds. To maximize the sum, we formulate a contrastive lower bound on the conditional MI which can be approximated efficiently. We refer to our general approach as Decomposed Estimation of Mutual Information (DEMI). We show that DEMI can capture a larger amount of MI than standard non-decomposed contrastive bounds in a synthetic setting, and learns better representations in a vision domain and for dialogue generation.
小貼士
登錄享
相關主題
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191