Search

While the activations of neurons in deep neural networks usually do not have a simple human-understandable interpretation, sparse autoencoders (SAEs) can be used to transform these activations into a higher-dimensional latent space which may be more easily interpretable. However, these SAEs can have millions of distinct latent features, making it infeasible for humans to manually interpret each one. In this work, we build an open-source automated pipeline to generate and evaluate natural language explanations for SAE features using LLMs. We test our framework on SAEs of varying sizes, activation functions, and losses, trained on two different open-weight LLMs. We introduce five new techniques to score the quality of explanations that are cheaper to run than the previous state of the art. One of these techniques, intervention scoring, evaluates the interpretability of the effects of intervening on a feature, which we find explains features that are not recalled by existing methods. We propose guidelines for generating better explanations that remain valid for a broader set of activating contexts, and discuss pitfalls with existing scoring techniques. We use our explanations to measure the semantic similarity of independently trained SAEs, and find that SAEs trained on nearby layers of the residual stream are highly similar. Our large-scale analysis confirms that SAE latents are indeed much more interpretable than neurons, even when neurons are sparsified using top-$k$ postprocessing. Our code is available at https://github.com/EleutherAI/sae-auto-interp, and our explanations are available at https://huggingface.co/datasets/EleutherAI/auto_interp_explanations.

In multi-agent learning, the predominant approach focuses on generalization, often neglecting the optimization of individual agents. This emphasis on generalization limits the ability of agents to utilize their unique strengths, resulting in inefficiencies. This paper introduces Comparative Advantage Maximization (CAM), a method designed to enhance individual agent specialization in multiagent systems. CAM employs a two-phase process, combining centralized population training with individual specialization through comparative advantage maximization. CAM achieved a 13.2% improvement in individual agent performance and a 14.9% increase in behavioral diversity compared to state-of-the-art systems. The success of CAM highlights the importance of individual agent specialization, suggesting new directions for multi-agent system development., Comment: 19 pages

Real-time dynamic scheduling is a crucial but notoriously challenging task in modern manufacturing processes due to its high decision complexity. Recently, reinforcement learning (RL) has been gaining attention as an impactful technique to handle this challenge. However, classical RL methods typically rely on human-made dispatching rules, which are not suitable for large-scale factory-wide scheduling. To bridge this gap, this paper applies a leader-follower multi-agent RL (MARL) concept to obtain desired coordination after decomposing the scheduling problem into a set of sub-problems that are handled by each individual agent for scalability. We further strengthen the procedure by proposing a rule-based conversion algorithm to prevent catastrophic loss of production capacity due to an agent's error. Our experimental results demonstrate that the proposed model outperforms the state-of-the-art deep RL-based scheduling models in various aspects. Additionally, the proposed model provides the most robust scheduling performance to demand changes. Overall, the proposed MARL-based scheduling model presents a promising solution to the real-time scheduling problem, with potential applications in various manufacturing industries.

Heavy precipitation from tropical cyclones (TCs) may result in disasters, such as floods and landslides, leading to substantial economic damage and loss of life. Prediction of TC precipitation based on ensemble post-processing procedures using machine learning (ML) approaches has received considerable attention for its flexibility in modeling and its computational power in managing complex models. However, when applying ML techniques to TC precipitation for a specific area, the available observation data are typically insufficient for comprehensive training, validation, and testing of the ML model, primarily due to the rapid movement of TCs. We propose to use the convolutional neural network (CNN) as a deep ML model to leverage the spatial information of precipitation. The proposed model has three distinct features that differentiate it from traditional CNNs applied in meteorology. First, it utilizes data augmentation to alleviate challenges posed by the small sample size. Second, it contains geographical and dynamic variables to account for area-specific features and the relative distance between the study area and the moving TC. Third, it applies unequal weights to accommodate the temporal structure in the training data when calculating the objective function. The proposed CNN-all model is then illustrated with the TC Soudelor's impact on Taiwan. Soudelor was the strongest TC of the 2015 Pacific typhoon season. The results show that the inclusion of augmented data and dynamic variables improves the prediction of heavy precipitation. The proposed CNN-all outperforms traditional CNN models, based on the continuous probability skill score (CRPSS), probability plots, and reliability diagram. The proposed model has the potential to be utilized in a wide range of meteorological studies., Comment: 21 pages, 6 figures, 2 tables, 2 supplementary figures

We introduce NNsight and NDIF, technologies that work in tandem to enable scientific study of very large neural networks. NNsight is an open-source system that extends PyTorch to introduce deferred remote execution. NDIF is a scalable inference service that executes NNsight requests, allowing users to share GPU resources and pretrained models. These technologies are enabled by the intervention graph, an architecture developed to decouple experiment design from model runtime. Together, this framework provides transparent and efficient access to the internals of deep neural networks such as very large language models (LLMs) without imposing the cost or complexity of hosting customized models individually. We conduct a quantitative survey of the machine learning literature that reveals a growing gap in the study of the internals of large-scale AI. We demonstrate the design and use of our framework to address this gap by enabling a range of research methods on huge models. Finally, we conduct benchmarks to compare performance with previous approaches. Code documentation, and materials are available at https://nnsight.net/., Comment: Code at https://nnsight.net

Task-oriented dialogue systems are broadly used in virtual assistants and other automated services, providing interfaces between users and machines to facilitate specific tasks. Nowadays, task-oriented dialogue systems have greatly benefited from pre-trained language models (PLMs). However, their task-solving performance is constrained by the inherent capacities of PLMs, and scaling these models is expensive and complex as the model size becomes larger. To address these challenges, we propose Soft Mixture-of-Expert Task-Oriented Dialogue system (SMETOD) which leverages an ensemble of Mixture-of-Experts (MoEs) to excel at subproblems and generate specialized outputs for task-oriented dialogues. SMETOD also scales up a task-oriented dialogue system with simplicity and flexibility while maintaining inference efficiency. We extensively evaluate our model on three benchmark functionalities: intent prediction, dialogue state tracking, and dialogue response generation. Experimental results demonstrate that SMETOD achieves state-of-the-art performance on most evaluated metrics. Moreover, comparisons against existing strong baselines show that SMETOD has a great advantage in the cost of inference and correctness in problem-solving.

Like a criminal under investigation, Large Language Models (LLMs) might pretend to be aligned while evaluated and misbehave when they have a good opportunity. Can current interpretability methods catch these 'alignment fakers?' To answer this question, we introduce a benchmark that consists of 324 pairs of LLMs fine-tuned to select actions in role-play scenarios. One model in each pair is consistently benign (aligned). The other model misbehaves in scenarios where it is unlikely to be caught (alignment faking). The task is to identify the alignment faking model using only inputs where the two models behave identically. We test five detection strategies, one of which identifies 98% of alignment-fakers.

The dual-channel multi-mode 850 nm optical Miniature Transmitter (MTx) is developed for data transmission of the ATLAS LAr calorimeter readout at LHC. The MTx's are exposed to the radiation field of proton-proton collisions, therefore, the tolerance in Total Ionizing Dose (TID) is required. The TID effects in the MTx are investigated with X-rays and Co-60 gamma-rays for the active components of VCSEL diodes and the customized Link-on-Chip laser driver (LOCld) developed in 0.25 um Silicon-on-Sapphire CMOS technology. The irradiation tests were conducted at various dose rates. The responses to TID are observed with degradation of laser currents at initial dose of 10 to 100 Gy(SiO2), and partial recovery with additional TID to a stable output about 90 % of the original. The optical eye diagrams of irradiated samples show slightly increased jittering, and are suitable for the ATLAS requirement of 5 Gbps applications., Comment: 7 pages, 10 figures

Tracking dialogue states is an essential topic in task-oriented dialogue systems, which involve filling in the necessary information in pre-defined slots corresponding to a schema. While general pre-trained language models have been shown effective in slot-filling, their performance is limited when applied to specific domains. We propose a graph-based framework that learns domain-specific prompts by incorporating the dialogue schema. Specifically, we embed domain-specific schema encoded by a graph neural network into the pre-trained language model, which allows for relations in the schema to guide the model for better adaptation to the specific domain. Our experiments demonstrate that the proposed graph-based method outperforms other multi-domain DST approaches while using similar or fewer trainable parameters. We also conduct a comprehensive study of schema graph architectures, parameter usage, and module ablation that demonstrate the effectiveness of our model on multi-domain dialogue state tracking.

Vector graphics are widely used in graphical designs and have received more and more attention. However, unlike raster images which can be easily obtained, acquiring high-quality vector graphics, typically through automatically converting from raster images remains a significant challenge, especially for more complex images such as photos or artworks. In this paper, we propose SAMVG, a multi-stage model to vectorize raster images into SVG (Scalable Vector Graphics). Firstly, SAMVG uses general image segmentation provided by the Segment-Anything Model and uses a novel filtering method to identify the best dense segmentation map for the entire image. Secondly, SAMVG then identifies missing components and adds more detailed components to the SVG. Through a series of extensive experiments, we demonstrate that SAMVG can produce high quality SVGs in any domain while requiring less computation time and complexity compared to previous state-of-the-art methods., Comment: Accepted by ICASSP 2024

This paper proposes a novel multi-agent reinforcement learning (MARL) method to learn multiple coordinated agents under directed acyclic graph (DAG) constraints. Unlike existing MARL approaches, our method explicitly exploits the DAG structure between agents to achieve more effective learning performance. Theoretically, we propose a novel surrogate value function based on a MARL model with synthetic rewards (MARLM-SR) and prove that it serves as a lower bound of the optimal value function. Computationally, we propose a practical training algorithm that exploits new notion of leader agent and reward generator and distributor agent to guide the decomposed follower agents to better explore the parameter space in environments with DAG constraints. Empirically, we exploit four DAG environments including a real-world scheduling for one of Intel's high volume packaging and test factory to benchmark our methods and show it outperforms the other non-DAG approaches.

With the demanding need for deploying dialogue systems in new domains with less cost, zero-shot dialogue state tracking (DST), which tracks user's requirements in task-oriented dialogues without training on desired domains, draws attention increasingly. Although prior works have leveraged question-answering (QA) data to reduce the need for in-domain training in DST, they fail to explicitly model knowledge transfer and fusion for tracking dialogue states. To address this issue, we propose CoFunDST, which is trained on domain-agnostic QA datasets and directly uses candidate choices of slot-values as knowledge for zero-shot dialogue-state generation, based on a T5 pre-trained language model. Specifically, CoFunDST selects highly-relevant choices to the reference context and fuses them to initialize the decoder to constrain the model outputs. Our experimental results show that our proposed model achieves outperformed joint goal accuracy compared to existing zero-shot DST approaches in most domains on the MultiWOZ 2.1. Extensive analyses demonstrate the effectiveness of our proposed approach for improving zero-shot DST learning from QA., Comment: Accepted by ICASSP 2023

Recent quasi-2D systems with judicious exploitation of the atomic monolayer or few-layer architecture exhibit unprecedented physical properties that challenge the conventional wisdom on the condensed matter physics. Here we show that the infinite layer SrCuO2 (SCO), a topical cuprate Mott insulator in the bulk form, can manifest an unexpected metallic state in the quasi-2D limit when SCO is grown on TiO2-terminated SrTiO3 (STO) substrates. Hard x-ray core-level photoemission spectra demonstrate a definitive Fermi level that resembles the hole doped metal. Soft x-ray absorption spectroscopy also reveals features analogous to those of a hole doped Mott insulator. Based on these results, we conclude that the hole doping does not occur at the interfaces between SCO and STO; instead, it comes from the transient layers between the chain type and the planar type structures within the SCO slab. The present work reveals a novel metallic state in the infinite layer SCO and invites further examination to elucidate the spatial extent of this state., Comment: 31 pages, 11 figures. Physical Review B, in press