In this work we present successor heads: attention heads that increment tokens with a natural ordering, such as numbers, months, and days. For example, successor heads increment 'Monday' into 'Tuesday'. We explain the successor head behavior with an approach rooted in mechanistic interpretability, the field that aims to explain how models complete tasks in human-understandable terms. Existing research in this area has found interpretable language model components in small toy models. However, results in toy models have not yet led to insights that explain the internals of frontier models and little is currently understood about the internal operations of large language models. In this paper, we analyze the behavior of successor heads in large language models (LLMs) and find that they implement abstract representations that are common to different architectures. They form in LLMs with as few as 31 million parameters, and at least as many as 12 billion parameters, such as GPT-2, Pythia, and Llama-2. We find a set of 'mod-10 features' that underlie how successor heads increment in LLMs across different architectures and sizes. We perform vector arithmetic with these features to edit head behavior and provide insights into numeric representations within LLMs. Additionally, we study the behavior of successor heads on natural language data, identifying interpretable polysemanticity in a Pythia successor head.
相關內容
ACM/IEEE第23屆模型驅動工程語言和系統國際會議，是模型驅動軟件和系統工程的首要會議系列，由ACM-SIGSOFT和IEEE-TCSE支持組織。自1998年以來，模型涵蓋了建模的各個方面，從語言和方法到工具和應用程序。模特的參加者來自不同的背景，包括研究人員、學者、工程師和工業專業人士。MODELS 2019是一個論壇，參與者可以圍繞建模和模型驅動的軟件和系統交流前沿研究成果和創新實踐經驗。今年的版本將為建模社區提供進一步推進建模基礎的機會，并在網絡物理系統、嵌入式系統、社會技術系統、云計算、大數據、機器學習、安全、開源等新興領域提出建模的創新應用以及可持續性。
官網鏈接： ·
Q函數
·
分解的
·
Markov
·
泛函
·
Many real-world applications involve some agents that fall into two teams, with payoffs that are equal within the same team but of opposite sign across the opponent team. The so-called two-team zero-sum Markov games (2t0sMGs) can be resolved with reinforcement learning in recent years. However, existing methods are thus inefficient in light of insufficient consideration of intra-team credit assignment, data utilization and computational intractability. In this paper, we propose the individual-global-minimax (IGMM) principle to ensure the coherence between two-team minimax behaviors and the individual greedy behaviors through Q functions in 2t0sMGs. Based on it, we present a novel multi-agent reinforcement learning framework, Factorized Multi-Agent MiniMax Q-Learning (FM3Q), which can factorize the joint minimax Q function into individual ones and iteratively solve for the IGMM-satisfied minimax Q functions for 2t0sMGs. Moreover, an online learning algorithm with neural networks is proposed to implement FM3Q and obtain the deterministic and decentralized minimax policies for two-team players. A theoretical analysis is provided to prove the convergence of FM3Q. Empirically, we use three environments to evaluate the learning efficiency and final performance of FM3Q and show its superiority on 2t0sMGs.
In recent years, more people have seen their work depend on data manipulation tasks. However, many of these users do not have the background in programming required to write complex programs, particularly SQL queries. One way of helping these users is automatically synthesizing the SQL query given a small set of examples. Several program synthesizers for SQL have been recently proposed, but they do not leverage multicore architectures. This paper proposes CUBES, a parallel program synthesizer for the domain of SQL queries using input-output examples. Since input-output examples are an under-specification of the desired SQL query, sometimes, the synthesized query does not match the user's intent. CUBES incorporates a new disambiguation procedure based on fuzzing techniques that interacts with the user and increases the confidence that the returned query matches the user intent. We perform an extensive evaluation on around 4000 SQL queries from different domains. Experimental results show that our sequential version can solve more instances than other state-of-the-art SQL synthesizers. Moreover, the parallel approach can scale up to 16 processes with super-linear speedups for many hard instances. Our disambiguation approach is critical to achieving an accuracy of around 60%, significantly larger than other SQL synthesizers.
Probability predictions are essential to inform decision making across many fields. Ideally, probability predictions are (i) well calibrated, (ii) accurate, and (iii) bold, i.e., spread out enough to be informative for decision making. However, there is a fundamental tension between calibration and boldness, since calibration metrics can be high when predictions are overly cautious, i.e., non-bold. The purpose of this work is to develop a Bayesian model selection-based approach to assess calibration, and a strategy for boldness-recalibration that enables practitioners to responsibly embolden predictions subject to their required level of calibration. Specifically, we allow the user to pre-specify their desired posterior probability of calibration, then maximally embolden predictions subject to this constraint. We demonstrate the method with a case study on hockey home team win probabilities and then verify the performance of our procedures via simulation. We find that very slight relaxation of calibration probability (e.g., from 0.99 to 0.95) can often substantially embolden predictions when they are well calibrated and accurate (e.g., widening hockey predictions range from .26-.78 to .10-.91).
As spiking neural networks receive more attention, we look toward applications of this computing paradigm in fields other than computer vision and signal processing. One major field, underexplored in the neuromorphic setting, is Natural Language Processing (NLP), where most state-of-the-art solutions still heavily rely on resource-consuming and power-hungry traditional deep learning architectures. Therefore, it is compelling to design NLP models for neuromorphic architectures due to their low energy requirements, with the additional benefit of a more human-brain-like operating model for processing information. However, one of the biggest issues with bringing NLP to the neuromorphic setting is in properly encoding text into a spike train so that it can be seamlessly handled by both current and future SNN architectures. In this paper, we compare various methods of encoding text as spikes and assess each method's performance in an associated SNN on a downstream NLP task, namely, sentiment analysis. Furthermore, we go on to propose a new method of encoding text as spikes that outperforms a widely-used rate-coding technique, Poisson rate-coding, by around 13\% on our benchmark NLP tasks. Subsequently, we demonstrate the energy efficiency of SNNs implemented in hardware for the sentiment analysis task compared to traditional deep neural networks, observing an energy efficiency increase of more than 32x during inference and 60x during training while incurring the expected energy-performance tradeoff.
Extensive fine-tuning on Large Language Models does not always yield better results. Oftentimes, models tend to get better at imitating one form of data without gaining greater reasoning ability and may even end up losing some intelligence. Here I introduce EvoMerge, a systematic approach to large language model training and merging. Leveraging model merging for weight crossover and fine-tuning for weight mutation, EvoMerge establishes an evolutionary process aimed at pushing models beyond the limits of conventional fine-tuning.
Data plays a fundamental role in the training of Large Language Models (LLMs). Effective data management, particularly in the formulation of a well-suited training dataset, holds significance for enhancing model performance and improving training efficiency during pretraining and supervised fine-tuning phases. Despite the considerable importance of data management, the current research community still falls short in providing a systematic analysis of the rationale behind management strategy selection, its consequential effects, methodologies for evaluating curated datasets, and the ongoing pursuit of improved strategies. Consequently, the exploration of data management has attracted more and more attention among the research community. This survey provides a comprehensive overview of current research in data management within both the pretraining and supervised fine-tuning stages of LLMs, covering various noteworthy aspects of data management strategy design: data quantity, data quality, domain/task composition, etc. Looking toward the future, we extrapolate existing challenges and outline promising directions for development in this field. Therefore, this survey serves as a guiding resource for practitioners aspiring to construct powerful LLMs through effective data management practices. The collection of the latest papers is available at //github.com/ZigeW/data_management_LLM.
Recent years have witnessed remarkable progress made in large language models (LLMs). Such advancements, while garnering significant attention, have concurrently elicited various concerns. The potential of these models is undeniably vast; however, they may yield texts that are imprecise, misleading, or even detrimental. Consequently, it becomes paramount to employ alignment techniques to ensure these models to exhibit behaviors consistent with human values. This survey endeavors to furnish an extensive exploration of alignment methodologies designed for LLMs, in conjunction with the extant capability research in this domain. Adopting the lens of AI alignment, we categorize the prevailing methods and emergent proposals for the alignment of LLMs into outer and inner alignment. We also probe into salient issues including the models' interpretability, and potential vulnerabilities to adversarial attacks. To assess LLM alignment, we present a wide variety of benchmarks and evaluation methodologies. After discussing the state of alignment research for LLMs, we finally cast a vision toward the future, contemplating the promising avenues of research that lie ahead. Our aspiration for this survey extends beyond merely spurring research interests in this realm. We also envision bridging the gap between the AI alignment research community and the researchers engrossed in the capability exploration of LLMs for both capable and safe LLMs.
We present CoDEx, a set of knowledge graph completion datasets extracted from Wikidata and Wikipedia that improve upon existing knowledge graph completion benchmarks in scope and level of difficulty. In terms of scope, CoDEx comprises three knowledge graphs varying in size and structure, multilingual descriptions of entities and relations, and tens of thousands of hard negative triples that are plausible but verified to be false. To characterize CoDEx, we contribute thorough empirical analyses and benchmarking experiments. First, we analyze each CoDEx dataset in terms of logical relation patterns. Next, we report baseline link prediction and triple classification results on CoDEx for five extensively tuned embedding models. Finally, we differentiate CoDEx from the popular FB15K-237 knowledge graph completion dataset by showing that CoDEx covers more diverse and interpretable content, and is a more difficult link prediction benchmark. Data, code, and pretrained models are available at //bit.ly/2EPbrJs.
The world we see is ever-changing and it always changes with people, things, and the environment. Domain is referred to as the state of the world at a certain moment. A research problem is characterized as domain transfer adaptation when it needs knowledge correspondence between different moments. Conventional machine learning aims to find a model with the minimum expected risk on test data by minimizing the regularized empirical risk on the training data, which, however, supposes that the training and test data share similar joint probability distribution. Transfer adaptation learning aims to build models that can perform tasks of target domain by learning knowledge from a semantic related but distribution different source domain. It is an energetic research filed of increasing influence and importance. This paper surveys the recent advances in transfer adaptation learning methodology and potential benchmarks. Broader challenges being faced by transfer adaptation learning researchers are identified, i.e., instance re-weighting adaptation, feature adaptation, classifier adaptation, deep network adaptation, and adversarial adaptation, which are beyond the early semi-supervised and unsupervised split. The survey provides researchers a framework for better understanding and identifying the research status, challenges and future directions of the field.
Explainable Recommendation refers to the personalized recommendation algorithms that address the problem of why -- they not only provide the user with the recommendations, but also make the user aware why such items are recommended by generating recommendation explanations, which help to improve the effectiveness, efficiency, persuasiveness, and user satisfaction of recommender systems. In recent years, a large number of explainable recommendation approaches -- especially model-based explainable recommendation algorithms -- have been proposed and adopted in real-world systems. In this survey, we review the work on explainable recommendation that has been published in or before the year of 2018. We first high-light the position of explainable recommendation in recommender system research by categorizing recommendation problems into the 5W, i.e., what, when, who, where, and why. We then conduct a comprehensive survey of explainable recommendation itself in terms of three aspects: 1) We provide a chronological research line of explanations in recommender systems, including the user study approaches in the early years, as well as the more recent model-based approaches. 2) We provide a taxonomy for explainable recommendation algorithms, including user-based, item-based, model-based, and post-model explanations. 3) We summarize the application of explainable recommendation in different recommendation tasks, including product recommendation, social recommendation, POI recommendation, etc. We devote a chapter to discuss the explanation perspectives in the broader IR and machine learning settings, as well as their relationship with explainable recommendation research. We end the survey by discussing potential future research directions to promote the explainable recommendation research area.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191