To mitigate memorization, we introduce a subtle modification to the next-token training objective that we call the goldfish loss. During training, a randomly sampled subset of tokens are excluded from the loss computation. These dropped tokens are not memorized by the model, which prevents verbatim reproduction of a complete chain of tokens from the training set. We run extensive experiments training billion-scale Llama-2 models, both pre-trained and trained from scratch, and demonstrate significant reductions in extractable memorization with little to no impact on downstream benchmarks.

Membership Inference Attacks (MIAs) determine whether a specific data point was included in the training set of a target model. In this paper, we introduce the Semantic Membership Inference Attack (SMIA), a novel approach that enhances MIA performance by leveraging the semantic content of inputs and their perturbations. SMIA trains a neural network to analyze the target model's behavior on perturbed inputs, effectively capturing variations in output probability distributions between members and non-members. We conduct comprehensive evaluations on the Pythia and GPT-Neo model families using the Wikipedia dataset. Our results show that SMIA significantly outperforms existing MIAs; for instance, SMIA achieves an AUC-ROC of 67.39% on Pythia-12B, compared to 58.90% by the second-best attack.

Enhancing Language Models' (LMs) ability to understand purchase intentions in E-commerce scenarios is crucial for their effective assistance in various downstream tasks. However, previous approaches that distill intentions from LMs often fail to generate meaningful and human-centric intentions applicable in real-world E-commerce contexts. This raises concerns about the true comprehension and utilization of purchase intentions by LMs. In this paper, we present IntentionQA, a double-task multiple-choice question answering benchmark to evaluate LMs' comprehension of purchase intentions in E-commerce. Specifically, LMs are tasked to infer intentions based on purchased products and utilize them to predict additional purchases. IntentionQA consists of 4,360 carefully curated problems across three difficulty levels, constructed using an automated pipeline to ensure scalability on large E-commerce platforms. Human evaluations demonstrate the high quality and low false-negative rate of our benchmark. Extensive experiments across 19 language models show that they still struggle with certain scenarios, such as understanding products and intentions accurately, jointly reasoning with products and intentions, and more, in which they fall far behind human performances. Our code and data are publicly available at https://github.com/HKUST-KnowComp/IntentionQA.

To address the need for robust positioning, navigation, and timing services in lunar and Martian environments, this paper proposes a novel fault detection framework for satellite constellations using inter-satellite ranging (ISR). Traditional fault monitoring methods rely on intense monitoring from ground-based stations, which are impractical for lunar and Martian missions due to cost constraints. Our approach leverages graph-rigidity theory to detect faults without relying on precise ephemeris. We model satellite constellations as graphs where satellites are vertices and inter-satellite links are edges. By analyzing the Euclidean Distance Matrix (EDM) derived from ISR measurements, we identify faults through the singular values of the geometric-centered EDM (GCEDM). A neural network predictor is employed to handle the diverse geometry of the graph, enhancing fault detection robustness. The proposed method is validated through simulations of constellations around Mars and the Moon, demonstrating its effectiveness in various configurations. This research contributes to the reliable operation of satellite constellations for future lunar and Martian exploration missions.

The incidence and mortality rates of malignant tumors, such as acute leukemia, have risen significantly. Clinically, hospitals rely on cytological examination of peripheral blood and bone marrow smears to diagnose malignant tumors, with accurate blood cell counting being crucial. Existing automated methods face challenges such as low feature expression capability, poor interpretability, and redundant feature extraction when processing high-dimensional microimage data. We propose a novel fine-grained classification model, SCKansformer, for bone marrow blood cells, which addresses these challenges and enhances classification accuracy and efficiency. The model integrates the Kansformer Encoder, SCConv Encoder, and Global-Local Attention Encoder. The Kansformer Encoder replaces the traditional MLP layer with the KAN, improving nonlinear feature representation and interpretability. The SCConv Encoder, with its Spatial and Channel Reconstruction Units, enhances feature representation and reduces redundancy. The Global-Local Attention Encoder combines Multi-head Self-Attention with a Local Part module to capture both global and local features. We validated our model using the Bone Marrow Blood Cell Fine-Grained Classification Dataset (BMCD-FGCD), comprising over 10,000 samples and nearly 40 classifications, developed with a partner hospital. Comparative experiments on our private dataset, as well as the publicly available PBC and ALL-IDB datasets, demonstrate that SCKansformer outperforms both typical and advanced microcell classification methods across all datasets. Our source code and private BMCD-FGCD dataset are available at https://github.com/JustlfC03/SCKansformer.

Despite recent surge of interest in deploying colon capsule endoscopy (CCE) for early diagnosis of colorectal diseases, there remains a large gap between the current state of CCE in clinical practice, and the state of its counterpart optical colonoscopy (OC). Our study is aimed at closing this gap, by focusing on the full integration of AI in CCE's pathway, where image processing steps linked to the detection, localization and characterisation of important findings are carried out autonomously using various AI algorithms. We developed a recognition network, that with an impressive sensitivity of 99.9%, a specificity of 99.4%, and a negative predictive value (NPV) of 99.8%, detected colorectal polyps. After recognising a polyp within a sequence of images, only those images containing polyps were fed into two parallel independent networks for characterisation, and estimation of the size of those important findings. The characterisation network reached a sensitivity of 82% and a specificity of 80% in classifying polyps to two groups, namely neoplastic vs. non-neoplastic. The size estimation network reached an accuracy of 88% in correctly segmenting the polyps. By automatically incorporating this crucial information into CCE's pathway, we moved a step closer towards the full integration of AI in CCE's routine clinical practice.

Large language models (LLMs) demonstrate impressive zero-shot and few-shot reasoning capabilities. Some propose that such capabilities can be improved through self-reflection, i.e., letting LLMs reflect on their own output to identify and correct mistakes in the initial responses. However, despite some evidence showing the benefits of self-reflection, recent studies offer mixed results. Here, we aim to reconcile these conflicting findings by first demonstrating that the outcome of self-reflection is sensitive to prompt wording; e.g., LLMs are more likely to conclude that it has made a mistake when explicitly prompted to find mistakes. Consequently, idiosyncrasies in reflection prompts may lead LLMs to change correct responses unnecessarily. We show that most prompts used in the self-reflection literature are prone to this bias. We then propose different ways of constructing prompts that are conservative in identifying mistakes and show that self-reflection using such prompts results in higher accuracy. Our findings highlight the importance of prompt engineering in self-reflection tasks. We release our code at https://github.com/Michael98Liu/mixture-of-prompts.

In the context of medical imaging and machine learning, one of the most pressing challenges is the effective adaptation of pre-trained models to specialized medical contexts. Despite the availability of advanced pre-trained models, their direct application to the highly specialized and diverse field of medical imaging often falls short due to the unique characteristics of medical data. This study provides a comprehensive analysis on the performance of various fine-tuning methods applied to pre-trained models across a spectrum of medical imaging domains, including X-ray, MRI, Histology, Dermoscopy, and Endoscopic surgery. We evaluated eight fine-tuning strategies, including standard techniques such as fine-tuning all layers or fine-tuning only the classifier layers, alongside methods such as gradually unfreezing layers, regularization based fine-tuning and adaptive learning rates. We selected three well-established CNN architectures (ResNet-50, DenseNet-121, and VGG-19) to cover a range of learning and feature extraction scenarios. Although our results indicate that the efficacy of these fine-tuning methods significantly varies depending on both the architecture and the medical imaging type, strategies such as combining Linear Probing with Full Fine-tuning resulted in notable improvements in over 50% of the evaluated cases, demonstrating general effectiveness across medical domains. Moreover, Auto-RGN, which dynamically adjusts learning rates, led to performance enhancements of up to 11% for specific modalities. Additionally, the DenseNet architecture showed more pronounced benefits from alternative fine-tuning approaches compared to traditional full fine-tuning. This work not only provides valuable insights for optimizing pre-trained models in medical image analysis but also suggests the potential for future research into more advanced architectures and fine-tuning methods.

Early disease detection in veterinary care relies on identifying subclinical abnormalities in asymptomatic animals during wellness visits. This study introduces an algorithm designed to distinguish between wellness and other veterinary visits.The purpose of this study is to validate the use of a visit classification algorithm compared to manual classification of veterinary visits by three board-certified veterinarians. Using a dataset of 11,105 clinical visits from 2012 to 2017 involving 655 animals (85.3% canines and 14.7% felines) across 544 U.S. veterinary establishments, the model was trained using a Gradient Boosting Machine model. Three validators were tasked with classifying 400 visits, including both wellness and other types of visits, selected randomly from the same database used for initial algorithm training, aiming to maintain consistency and relevance between the training and application phases; visit classifications were subsequently categorized into "wellness" or "other" based on majority consensus among validators to assess the algorithm's performance in identifying wellness visits. The algorithm demonstrated a specificity of 0.94 (95% CI: 0.91 to 0.96), implying its accuracy in distinguishing non-wellness visits. The algorithm had a sensitivity of 0.86 (95% CI: 0.80 to 0.92), indicating its ability to correctly identify wellness visits as compared to the annotations provided by veterinary experts. The balanced accuracy, calculated as 0.90 (95% CI: 0.87 to 0.93), further confirms the algorithm's overall effectiveness. The algorithm exhibits strong specificity and sensitivity, ensuring accurate identification of a high proportion of wellness visits. Overall, this algorithm holds promise for advancing research on preventive care's role in subclinical disease identification, but prospective studies are needed for validation.

Granger causality, commonly used for inferring causal structures from time series data, has been adopted in widespread applications across various fields due to its intuitive explainability and high compatibility with emerging deep neural network prediction models. To alleviate challenges in better deciphering causal structures unambiguously from time series, the use of interventional data has become a practical approach. However, existing methods have yet to be explored in the context of imperfect interventions with unknown targets, which are more common and often more beneficial in a wide range of real-world applications. Additionally, the identifiability issues of Granger causality with unknown interventional targets in complex network models remain unsolved. Our work presents a theoretically-grounded method that infers Granger causal structure and identifies unknown targets by leveraging heterogeneous interventional time series data. We further illustrate that learning Granger causal structure and recovering interventional targets can mutually promote each other. Comparative experiments demonstrate that our method outperforms several robust baseline methods in learning Granger causal structure from interventional time series data.