Foundations models are presented as generalists that often perform well over a myriad of tasks. Fine-tuning these models, even on limited data, provides an additional boost in task-specific performance but often at the cost of their wider generalization, an effect termed catastrophic forgetting. In this paper, we analyze the relation between task difficulty in the CLIP model and the performance of several simple parameter-efficient fine-tuning methods through the lens of domain generalization and catastrophic forgetting. We provide evidence that the silhouette score of the zero-shot image and text embeddings is a better measure of task difficulty than the average cosine similarity of correct image/label embeddings, and discuss observable relationships between task difficulty, fine-tuning method, domain generalization, and catastrophic forgetting. Additionally, the averaged results across tasks and performance measures demonstrate that a simplified method that trains only a subset of attention weights, which we call A-CLIP, yields a balance between domain generalization and catastrophic forgetting.
相關內容
The popularity of Android means it is a common target for malware. Over the years, various studies have found that machine learning models can effectively discriminate malware from benign applications. However, as the operating system evolves, so does malware, bringing into question the findings of these previous studies, many of which report very high accuracies using small, outdated, and often imbalanced datasets. In this paper, we reimplement 18 representative past works and reevaluate them using a balanced, relevant, and up-to-date dataset comprising 124,000 applications. We also carry out new experiments designed to fill holes in existing knowledge, and use our findings to identify the most effective features and models to use for Android malware detection within a contemporary environment. We show that high detection accuracies (up to 96.8%) can be achieved using features extracted through static analysis alone, yielding a modest benefit (1%) from using far more expensive dynamic analysis. API calls and opcodes are the most productive static and TCP network traffic provide the most predictive dynamic features. Random forests are generally the most effective model, outperforming more complex deep learning approaches. Whilst directly combining static and dynamic features is generally ineffective, ensembling models separately leads to performances comparable to the best models but using less brittle features.
In collaborative goal-oriented settings, the participants are not only interested in achieving a successful outcome, but do also implicitly negotiate the effort they put into the interaction (by adapting to each other). In this work, we propose a challenging interactive reference game that requires two players to coordinate on vision and language observations. The learning signal in this game is a score (given after playing) that takes into account the achieved goal and the players' assumed efforts during the interaction. We show that a standard Proximal Policy Optimization (PPO) setup achieves a high success rate when bootstrapped with heuristic partner behaviors that implement insights from the analysis of human-human interactions. And we find that a pairing of neural partners indeed reduces the measured joint effort when playing together repeatedly. However, we observe that in comparison to a reasonable heuristic pairing there is still room for improvement -- which invites further research in the direction of cost-sharing in collaborative interactions.
Artificial Intelligence (AI) models are now being utilized in all facets of our lives such as healthcare, education and employment. Since they are used in numerous sensitive environments and make decisions that can be life altering, potential biased outcomes are a pressing matter. Developers should ensure that such models don't manifest any unexpected discriminatory practices like partiality for certain genders, ethnicities or disabled people. With the ubiquitous dissemination of AI systems, researchers and practitioners are becoming more aware of unfair models and are bound to mitigate bias in them. Significant research has been conducted in addressing such issues to ensure models don't intentionally or unintentionally perpetuate bias. This survey offers a synopsis of the different ways researchers have promoted fairness in AI systems. We explore the different definitions of fairness existing in the current literature. We create a comprehensive taxonomy by categorizing different types of bias and investigate cases of biased AI in different application domains. A thorough study is conducted of the approaches and techniques employed by researchers to mitigate bias in AI models. Moreover, we also delve into the impact of biased models on user experience and the ethical considerations to contemplate when developing and deploying such models. We hope this survey helps researchers and practitioners understand the intricate details of fairness and bias in AI systems. By sharing this thorough survey, we aim to promote additional discourse in the domain of equitable and responsible AI.
Multimodal Large Language Models (MLLMs) have showcased impressive skills in tasks related to visual understanding and reasoning. Yet, their widespread application faces obstacles due to the high computational demands during both the training and inference phases, restricting their use to a limited audience within the research and user communities. In this paper, we investigate the design aspects of Multimodal Small Language Models (MSLMs) and propose an efficient multimodal assistant named Mipha, which is designed to create synergy among various aspects: visual representation, language models, and optimization strategies. We show that without increasing the volume of training data, our Mipha-3B outperforms the state-of-the-art large MLLMs, especially LLaVA-1.5-13B, on multiple benchmarks. Through detailed discussion, we provide insights and guidelines for developing strong MSLMs that rival the capabilities of MLLMs. Our code is available at //github.com/zhuyiche/llava-phi.
The rise of large language models (LLMs) has opened new opportunities in Recommender Systems (RSs) by enhancing user behavior modeling and content understanding. However, current approaches that integrate LLMs into RSs solely utilize either LLM or conventional recommender model (CRM) to generate final recommendations, without considering which data segments LLM or CRM excel in. To fill in this gap, we conduct experiments on MovieLens-1M and Amazon-Books datasets, and compare the performance of a representative CRM (DCNv2) and an LLM (LLaMA2-7B) on various groups of data samples. Our findings reveal that LLMs excel in data segments where CRMs exhibit lower confidence and precision, while samples where CRM excels are relatively challenging for LLM, requiring substantial training data and a long training time for comparable performance. This suggests potential synergies in the combination between LLM and CRM. Motivated by these insights, we propose Collaborative Recommendation with conventional Recommender and Large Language Model (dubbed \textit{CoReLLa}). In this framework, we first jointly train LLM and CRM and address the issue of decision boundary shifts through alignment loss. Then, the resource-efficient CRM, with a shorter inference time, handles simple and moderate samples, while LLM processes the small subset of challenging samples for CRM. Our experimental results demonstrate that CoReLLa outperforms state-of-the-art CRM and LLM methods significantly, underscoring its effectiveness in recommendation tasks.
Language models have shown promise in various tasks but can be affected by undesired data during training, fine-tuning, or alignment. For example, if some unsafe conversations are wrongly annotated as safe ones, the model fine-tuned on these samples may be harmful. Therefore, the correctness of annotations, i.e., the credibility of the dataset, is important. This study focuses on the credibility of real-world datasets, including the popular benchmarks Jigsaw Civil Comments, Anthropic Harmless & Red Team, PKU BeaverTails & SafeRLHF, that can be used for training a harmless language model. Given the cost and difficulty of cleaning these datasets by humans, we introduce a systematic framework for evaluating the credibility of datasets, identifying label errors, and evaluating the influence of noisy labels in the curated language data, specifically focusing on unsafe comments and conversation classification. With the framework, we find and fix an average of 6.16% label errors in 11 datasets constructed from the above benchmarks. The data credibility and downstream learning performance can be remarkably improved by directly fixing label errors, indicating the significance of cleaning existing real-world datasets. We provide an open-source tool, Docta, for data cleaning at //github.com/Docta-ai/docta.
Large language models have catalyzed an unprecedented wave in code generation. While achieving significant advances, they blur the distinctions between machine- and human-authored source code, causing integrity and authenticity issues of software artifacts. Previous methods such as DetectGPT have proven effective in discerning machine-generated texts, but they do not identify and harness the unique patterns of machine-generated code. Thus, its applicability falters when applied to code. In this paper, we carefully study the specific patterns that characterize machine- and human-authored code. Through a rigorous analysis of code attributes such as lexical diversity, conciseness, and naturalness, we expose unique patterns inherent to each source. We particularly notice that the syntactic segmentation of code is a critical factor in identifying its provenance. Based on our findings, we propose DetectCodeGPT, a novel method for detecting machine-generated code, which improves DetectGPT by capturing the distinct stylized patterns of code. Diverging from conventional techniques that depend on external LLMs for perturbations, DetectCodeGPT perturbs the code corpus by strategically inserting spaces and newlines, ensuring both efficacy and efficiency. Experiment results show that our approach significantly outperforms state-of-the-art techniques in detecting machine-generated code.
Researchers have proposed to use data of human preference feedback to fine-tune text-to-image generative models. However, the scalability of human feedback collection has been limited by its reliance on manual annotation. Therefore, we develop and test a method to automatically annotate user preferences from their spontaneous facial expression reaction to the generated images. We collect a dataset of Facial Expression Reaction to Generated Images (FERGI) and show that the activations of multiple facial action units (AUs) are highly correlated with user evaluations of the generated images. Specifically, AU4 (brow lowerer) is reflective of negative evaluations of the generated image whereas AU12 (lip corner puller) is reflective of positive evaluations. These can be useful in two ways. Firstly, we can automatically annotate user preferences between image pairs with substantial difference in these AU responses with an accuracy significantly outperforming state-of-the-art scoring models. Secondly, directly integrating the AU responses with the scoring models improves their consistency with human preferences. Finally, this method of automatic annotation with facial expression analysis can be potentially generalized to other generation tasks. The code is available at //github.com/ShuangquanFeng/FERGI, and the dataset is also available at the same link for research purposes.
Big models have achieved revolutionary breakthroughs in the field of AI, but they might also pose potential concerns. Addressing such concerns, alignment technologies were introduced to make these models conform to human preferences and values. Despite considerable advancements in the past year, various challenges lie in establishing the optimal alignment strategy, such as data cost and scalable oversight, and how to align remains an open question. In this survey paper, we comprehensively investigate value alignment approaches. We first unpack the historical context of alignment tracing back to the 1920s (where it comes from), then delve into the mathematical essence of alignment (what it is), shedding light on the inherent challenges. Following this foundation, we provide a detailed examination of existing alignment methods, which fall into three categories: Reinforcement Learning, Supervised Fine-Tuning, and In-context Learning, and demonstrate their intrinsic connections, strengths, and limitations, helping readers better understand this research area. In addition, two emerging topics, personal alignment, and multimodal alignment, are also discussed as novel frontiers in this field. Looking forward, we discuss potential alignment paradigms and how they could handle remaining challenges, prospecting where future alignment will go.
Diffusion models are a class of deep generative models that have shown impressive results on various tasks with dense theoretical founding. Although diffusion models have achieved impressive quality and diversity of sample synthesis than other state-of-the-art models, they still suffer from costly sampling procedure and sub-optimal likelihood estimation. Recent studies have shown great enthusiasm on improving the performance of diffusion model. In this article, we present a first comprehensive review of existing variants of the diffusion models. Specifically, we provide a first taxonomy of diffusion models and categorize them variants to three types, namely sampling-acceleration enhancement, likelihood-maximization enhancement and data-generalization enhancement. We also introduce in detail other five generative models (i.e., variational autoencoders, generative adversarial networks, normalizing flow, autoregressive models, and energy-based models), and clarify the connections between diffusion models and these generative models. Then we make a thorough investigation into the applications of diffusion models, including computer vision, natural language processing, waveform signal processing, multi-modal modeling, molecular graph generation, time series modeling, and adversarial purification. Furthermore, we propose new perspectives pertaining to the development of this generative model.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191