We introduce methods for discovering and applying sparse feature circuits. These are causally implicated subnetworks of human-interpretable features for explaining language model behaviors. Circuits identified in prior work consist of polysemantic and difficult-to-interpret units like attention heads or neurons, rendering them unsuitable for many downstream applications. In contrast, sparse feature circuits enable detailed understanding of unanticipated mechanisms. Because they are based on fine-grained units, sparse feature circuits are useful for downstream tasks: We introduce SHIFT, where we improve the generalization of a classifier by ablating features that a human judges to be task-irrelevant. Finally, we demonstrate an entirely unsupervised and scalable interpretability pipeline by discovering thousands of sparse feature circuits for automatically discovered model behaviors.
相關內容
Language has been useful in extending the vision encoder to data from diverse distributions without empirical discovery in training domains. However, as the image description is mostly at coarse-grained level and ignores visual details, the resulted embeddings are still ineffective in overcoming complexity of domains at inference time. We present a self-supervision framework WIDIn, Wording Images for Domain-Invariant representation, to disentangle discriminative visual representation, by only leveraging data in a single domain and without any test prior. Specifically, for each image, we first estimate the language embedding with fine-grained alignment, which can be consequently used to adaptively identify and then remove domain-specific counterpart from the raw visual embedding. WIDIn can be applied to both pretrained vision-language models like CLIP, and separately trained uni-modal models like MoCo and BERT. Experimental studies on three domain generalization datasets demonstrate the effectiveness of our approach.
In the era of artificial intelligence, the diversity of data modalities and annotation formats often renders data unusable directly, requiring understanding and format conversion before it can be used by researchers or developers with different needs. To tackle this problem, this article introduces a framework called Dataset Description Language (DSDL) that aims to simplify dataset processing by providing a unified standard for AI datasets. DSDL adheres to the three basic practical principles of generic, portable, and extensible, using a unified standard to express data of different modalities and structures, facilitating the dissemination of AI data, and easily extending to new modalities and tasks. The standardized specifications of DSDL reduce the workload for users in data dissemination, processing, and usage. To further improve user convenience, we provide predefined DSDL templates for various tasks, convert mainstream datasets to comply with DSDL specifications, and provide comprehensive documentation and DSDL tools. These efforts aim to simplify the use of AI data, thereby improving the efficiency of AI development.
The widespread use of mobile devices for all kind of transactions makes necessary reliable and real-time identity authentication, leading to the adoption of face recognition (FR) via the cameras embedded in such devices. Progress of deep Convolutional Neural Networks (CNNs) has provided substantial advances in FR. Nonetheless, the size of state-of-the-art architectures is unsuitable for mobile deployment, since they often encompass hundreds of megabytes and millions of parameters. We address this by studying methods for deep network compression applied to FR. In particular, we apply network pruning based on Taylor scores, where less important filters are removed iteratively. The method is tested on three networks based on the small SqueezeNet (1.24M parameters) and the popular MobileNetv2 (3.5M) and ResNet50 (23.5M) architectures. These have been selected to showcase the method on CNNs with different complexities and sizes. We observe that a substantial percentage of filters can be removed with minimal performance loss. Also, filters with the highest amount of output channels tend to be removed first, suggesting that high-dimensional spaces within popular CNNs are over-dimensionated.
The drive to develop artificial neural networks that efficiently utilize resources has generated significant interest in bio-inspired Spiking Neural Networks (SNNs). These networks are particularly attractive due to their potential in applications requiring low power and memory. This potential is further enhanced by the ability to perform online local learning, enabling them to adapt to dynamic environments. This requires the model to be adaptive in a self-supervised manner. While self-supervised learning has seen great success in many deep learning domains, its application for online local learning in multi-layer SNNs remains underexplored. In this paper, we introduce the "EchoSpike Predictive Plasticity" (ESPP) learning rule, a pioneering online local learning rule designed to leverage hierarchical temporal dynamics in SNNs through predictive and contrastive coding. We validate the effectiveness of this approach using benchmark datasets, demonstrating that it performs on par with current state-of-the-art supervised learning rules. The temporal and spatial locality of ESPP makes it particularly well-suited for low-cost neuromorphic processors, representing a significant advancement in developing biologically plausible self-supervised learning models for neuromorphic computing at the edge.
Combining semantic information with behavioral data is a crucial research area in recommender systems. A promising approach involves leveraging external knowledge to enrich behavioral-based recommender systems with abundant semantic information. However, this approach faces two primary challenges: denoising raw external knowledge and adapting semantic representations. To address these challenges, we propose an External Knowledge-Enhanced Recommendation method with LLM Assistance (TRAWL). This method utilizes large language models (LLMs) to extract relevant recommendation knowledge from raw external data and employs a contrastive learning strategy for adapter training. Experiments on public datasets and real-world online recommender systems validate the effectiveness of our approach.
Recent knowledge editing methods have primarily focused on modifying structured knowledge in large language models, heavily relying on the assumption that structured knowledge is stored as key-value pairs locally in MLP layers or specific neurons. However, this task setting overlooks the fact that a significant portion of real-world knowledge is stored in an unstructured format, characterized by long-form content, noise, and a complex yet comprehensive nature. The "knowledge locating" and "term-driven optimization" techniques conducted from the assumption used in previous methods (e.g., MEMIT) are ill-suited for unstructured knowledge. To address these challenges, we propose a novel unstructured knowledge editing method, namely UnKE, which extends previous assumptions in the layer dimension and token dimension. Firstly, in the layer dimension, we discard the "knowledge locating" step and treat first few layers as the key, which expand knowledge storage through layers to break the "knowledge stored locally" assumption. Next, we replace "term-driven optimization" with "cause-driven optimization" across all inputted tokens in the token dimension, directly optimizing the last layer of the key generator to perform editing to generate the required key vectors. By utilizing key-value pairs at the layer level, UnKE effectively represents and edits complex and comprehensive unstructured knowledge, leveraging the potential of both the MLP and attention layers. Results on newly proposed unstructure knowledge editing dataset (UnKEBench) and traditional structured datasets demonstrate that UnKE achieves remarkable performance, surpassing strong baselines.
We propose the Data Contamination Quiz (DCQ), a simple and effective approach to detect data contamination in large language models (LLMs) and estimate the amount of it. Specifically, we frame data contamination detection as a series of multiple-choice questions and devise a quiz format wherein three perturbed versions of each subsampled instance from a specific dataset partition (e.g., GSM8k test set) are created. These changes only include word-level perturbations. The generated perturbations, along with the original dataset instance, form the options in the DCQ, with an extra option accommodating the possibility of selecting none of the provided options. Given that the only distinguishing signal among the options is the exact wording with respect to the original dataset instance, an LLM, when tasked with identifying the original dataset instance, gravitates towards selecting the original one if it has been exposed to it in its pre-training phase -- a trait intrinsic to LLMs. While accounting for positional biases in LLMs, the quiz performance reveals the contamination level for the model being examined with the dataset partition to which the quiz pertains. Applied to various datasets with GPT-4 and GPT-3.5, our findings -- while fully lacking access to pre-training data and model parameters -- suggest that DCQ achieves state-of-the-art results and uncovers greater contamination/memorization levels compared to existing methods and proficiently bypasses more safety filters, especially those set to avoid generating copyrighted contents.
Believable proxies of human behavior can empower interactive applications ranging from immersive environments to rehearsal spaces for interpersonal communication to prototyping tools. In this paper, we introduce generative agents--computational software agents that simulate believable human behavior. Generative agents wake up, cook breakfast, and head to work; artists paint, while authors write; they form opinions, notice each other, and initiate conversations; they remember and reflect on days past as they plan the next day. To enable generative agents, we describe an architecture that extends a large language model to store a complete record of the agent's experiences using natural language, synthesize those memories over time into higher-level reflections, and retrieve them dynamically to plan behavior. We instantiate generative agents to populate an interactive sandbox environment inspired by The Sims, where end users can interact with a small town of twenty five agents using natural language. In an evaluation, these generative agents produce believable individual and emergent social behaviors: for example, starting with only a single user-specified notion that one agent wants to throw a Valentine's Day party, the agents autonomously spread invitations to the party over the next two days, make new acquaintances, ask each other out on dates to the party, and coordinate to show up for the party together at the right time. We demonstrate through ablation that the components of our agent architecture--observation, planning, and reflection--each contribute critically to the believability of agent behavior. By fusing large language models with computational, interactive agents, this work introduces architectural and interaction patterns for enabling believable simulations of human behavior.
Inspired by the human cognitive system, attention is a mechanism that imitates the human cognitive awareness about specific information, amplifying critical details to focus more on the essential aspects of data. Deep learning has employed attention to boost performance for many applications. Interestingly, the same attention design can suit processing different data modalities and can easily be incorporated into large networks. Furthermore, multiple complementary attention mechanisms can be incorporated in one network. Hence, attention techniques have become extremely attractive. However, the literature lacks a comprehensive survey specific to attention techniques to guide researchers in employing attention in their deep models. Note that, besides being demanding in terms of training data and computational resources, transformers only cover a single category in self-attention out of the many categories available. We fill this gap and provide an in-depth survey of 50 attention techniques categorizing them by their most prominent features. We initiate our discussion by introducing the fundamental concepts behind the success of attention mechanism. Next, we furnish some essentials such as the strengths and limitations of each attention category, describe their fundamental building blocks, basic formulations with primary usage, and applications specifically for computer vision. We also discuss the challenges and open questions related to attention mechanism in general. Finally, we recommend possible future research directions for deep attention.
With the advent of deep neural networks, learning-based approaches for 3D reconstruction have gained popularity. However, unlike for images, in 3D there is no canonical representation which is both computationally and memory efficient yet allows for representing high-resolution geometry of arbitrary topology. Many of the state-of-the-art learning-based 3D reconstruction approaches can hence only represent very coarse 3D geometry or are limited to a restricted domain. In this paper, we propose occupancy networks, a new representation for learning-based 3D reconstruction methods. Occupancy networks implicitly represent the 3D surface as the continuous decision boundary of a deep neural network classifier. In contrast to existing approaches, our representation encodes a description of the 3D output at infinite resolution without excessive memory footprint. We validate that our representation can efficiently encode 3D structure and can be inferred from various kinds of input. Our experiments demonstrate competitive results, both qualitatively and quantitatively, for the challenging tasks of 3D reconstruction from single images, noisy point clouds and coarse discrete voxel grids. We believe that occupancy networks will become a useful tool in a wide variety of learning-based 3D tasks.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191