Large language models (LLMs) have demonstrated remarkable prediction performance for a growing array of tasks. However, their rapid proliferation and increasing opaqueness have created a growing need for interpretability. Here, we ask whether we can automatically obtain natural language explanations for black box text modules. A "text module" is any function that maps text to a scalar continuous value, such as a submodule within an LLM or a fitted model of a brain region. "Black box" indicates that we only have access to the module's inputs/outputs. We introduce Summarize and Score (SASC), a method that takes in a text module and returns a natural language explanation of the module's selectivity along with a score for how reliable the explanation is. We study SASC in 3 contexts. First, we evaluate SASC on synthetic modules and find that it often recovers ground truth explanations. Second, we use SASC to explain modules found within a pre-trained BERT model, enabling inspection of the model's internals. Finally, we show that SASC can generate explanations for the response of individual fMRI voxels to language stimuli, with potential applications to fine-grained brain mapping. All code for using SASC and reproducing results is made available on Github.
相關內容
Recently, there has been a surge in the popularity of pre trained large language models (LLMs) (such as GPT-4), sweeping across the entire Natural Language Processing (NLP) and Computer Vision (CV) communities. These LLMs have demonstrated advanced multi-modal understanding capabilities and showcased strong performance across various benchmarks. The LLM has started to embody traits of artificial general intelligence, which holds vital guidance for enhancing brain-like characteristics within visual encoding models. Hence, This paper proposes a new multi-modal training paradigm, aligning with LLM, for encoding fMRI activity in visual cortex. Based on this paradigm, we trained an encoding model in fMRI data named the LLM-Visual Encoding Model (LLM-VEM). Specifically, we utilize LLM (miniGPT4) to generate descriptive text for all stimulus images, forming a high-quality textual description set. Moreover, we use the pre-trained text encoder (CLIP) to process these detailed descriptions, obtaining the text embedding features. Next, we use the contrast loss function to minimize the distance between the image embedding features and the text embedding features to complete the alignment operation of the stimulus image and text information. With the assistance of the pre-trained LLM, this alignment process facilitates better learning of the visual encoding model, resulting in higher precision. The final experimental results indicate that our training paradigm has significantly aided in enhancing the performance of the visual encoding model.
A new variant of Newton's method - named Backtracking New Q-Newton's method (BNQN) - which has strong theoretical guarantee, is easy to implement, and has good experimental performance, was recently introduced by the third author. Experiments performed previously showed some remarkable properties of the basins of attractions for finding roots of polynomials and meromorphic functions, with BNQN. In general, they look more smooth than that of Newton's method. In this paper, we continue to experimentally explore in depth this remarkable phenomenon, and connect BNQN to Newton's flow and Voronoi's diagram. This link poses a couple of challenging puzzles to be explained. Experiments also indicate that BNQN is more robust against random perturbations than Newton's method and Random Relaxed Newton's method.
Human trajectory forecasting is a critical challenge in fields such as robotics and autonomous driving. Due to the inherent uncertainty of human actions and intentions in real-world scenarios, various unexpected occurrences may arise. To uncover latent motion patterns in human behavior, we introduce a novel memory-based method, named Motion Pattern Priors Memory Network. Our method involves constructing a memory bank derived from clustered prior knowledge of motion patterns observed in the training set trajectories. We introduce an addressing mechanism to retrieve the matched pattern and the potential target distributions for each prediction from the memory bank, which enables the identification and retrieval of natural motion patterns exhibited by agents, subsequently using the target priors memory token to guide the diffusion model to generate predictions. Extensive experiments validate the effectiveness of our approach, achieving state-of-the-art trajectory prediction accuracy. The code will be made publicly available.
Two-sample spiked model is an important issue in multivariate statistical inference. This paper focuses on testing the number of spikes in a high-dimensional generalized two-sample spiked model, which is free of Gaussian population assumption and the diagonal or block-wise diagonal restriction of population covariance matrix, and the spiked eigenvalues are not necessary required to be bounded. In order to determine the number of spikes, we first propose a general test, which relies on the partial linear spectral statistics. We establish its asymptotic normality under the null hypothesis. Then we apply the conclusion to two statistical problem, variable selection in large-dimensional linear regression and change point detection when change points and additive outliers exist simultaneously. Simulations and empirical analysis are conducted to illustrate the good performance of our methods.
Multiple-conclusion Hilbert-style systems allow us to finitely axiomatize every logic defined by a finite matrix. Having obtained such axiomatizations for Paraconsistent Weak Kleene and Bochvar-Kleene logics, we modify them by replacing the multiple-conclusion rules with carefully selected single-conclusion ones. In this way we manage to introduce the first finite Hilbert-style single-conclusion axiomatizations for these logics.
The proximal gradient method is a generic technique introduced to tackle the non-smoothness in optimization problems, wherein the objective function is expressed as the sum of a differentiable convex part and a non-differentiable regularization term. Such problems with tensor format are of interest in many fields of applied mathematics such as image and video processing. Our goal in this paper is to address the solution of such problems with a more general form of the regularization term. An adapted iterative proximal gradient method is introduced for this purpose. Due to the slowness of the proposed algorithm, we use new tensor extrapolation methods to enhance its convergence. Numerical experiments on color image deblurring are conducted to illustrate the efficiency of our approach.
Graph Neural Networks (GNNs) are a class of deep learning models capable of processing graph-structured data, and they have demonstrated significant performance in a variety of real-world applications. Recent studies have found that GNN models are vulnerable to backdoor attacks. When specific patterns (called backdoor triggers, e.g., subgraphs, nodes, etc.) appear in the input data, the backdoor embedded in the GNN models is activated, which misclassifies the input data into the target class label specified by the attacker, whereas when there are no backdoor triggers in the input, the backdoor embedded in the GNN models is not activated, and the models work normally. Backdoor attacks are highly stealthy and expose GNN models to serious security risks. Currently, research on backdoor attacks against GNNs mainly focus on tasks such as graph classification and node classification, and backdoor attacks against link prediction tasks are rarely studied. In this paper, we propose a backdoor attack against the link prediction tasks based on GNNs and reveal the existence of such security vulnerability in GNN models, which make the backdoored GNN models to incorrectly predict unlinked two nodes as having a link relationship when a trigger appear. The method uses a single node as the trigger and poison selected node pairs in the training graph, and then the backdoor will be embedded in the GNN models through the training process. In the inference stage, the backdoor in the GNN models can be activated by simply linking the trigger node to the two end nodes of the unlinked node pairs in the input data, causing the GNN models to produce incorrect link prediction results for the target node pairs.
The present work is devoted to strong approximations of a generalized Ait-Sahalia model arising from mathematical finance. The numerical study of the considered model faces essential difficulties caused by a drift that blows up at the origin, highly nonlinear drift and diffusion coefficients and positivity-preserving requirement. In this paper, a novel explicit Euler-type scheme is proposed, which is easily implementable and able to preserve positivity of the original model unconditionally, i.e., for any time step-size h>0. A mean-square convergence rate of order 0.5 is also obtained for the proposed scheme in both non-critical and general critical cases. Our work is motivated by the need to justify the multi-level Monte Carlo (MLMC) simulations for the underlying model, where the rate of mean-square convergence is required and the preservation of positivity is desirable particularly for large discretization time steps. To the best of our knowledge, this is the first paper to propose an unconditionally positivity preserving explicit scheme with order 1/2 of mean-square convergence for the model. Numerical experiments are finally provided to confirm the theoretical findings.
This work introduces UstanceBR, a multimodal corpus in the Brazilian Portuguese Twitter domain for target-based stance prediction. The corpus comprises 86.8 k labelled stances towards selected target topics, and extensive network information about the users who published these stances on social media. In this article we describe the corpus multimodal data, and a number of usage examples in both in-domain and zero-shot stance prediction based on text- and network-related information, which are intended to provide initial baseline results for future studies in the field.
We present ResMLP, an architecture built entirely upon multi-layer perceptrons for image classification. It is a simple residual network that alternates (i) a linear layer in which image patches interact, independently and identically across channels, and (ii) a two-layer feed-forward network in which channels interact independently per patch. When trained with a modern training strategy using heavy data-augmentation and optionally distillation, it attains surprisingly good accuracy/complexity trade-offs on ImageNet. We will share our code based on the Timm library and pre-trained models.
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