Natural language supervision has been shown to be effective for zero-shot learning in many computer vision tasks, such as object detection and activity recognition. However, generating informative prompts can be challenging for more subtle tasks, such as video content moderation. This can be difficult, as there are many reasons why a video might be inappropriate, beyond violence and obscenity. For example, scammers may attempt to create junk content that is similar to popular educational videos but with no meaningful information. This paper evaluates the performance of several CLIP variations for content moderation of children's cartoons in both the supervised and zero-shot setting. We show that our proposed model (Vanilla CLIP with Projection Layer) outperforms previous work conducted on the Malicious or Benign (MOB) benchmark for video content moderation. This paper presents an in depth analysis of how context-specific language prompts affect content moderation performance. Our results indicate that it is important to include more context in content moderation prompts, particularly for cartoon videos as they are not well represented in the CLIP training data.

Machine learning models are often trained to predict the outcome resulting from a human decision. For example, if a doctor decides to test a patient for disease, will the patient test positive? A challenge is that the human decision censors the outcome data: we only observe test outcomes for patients doctors historically tested. Untested patients, for whom outcomes are unobserved, may differ from tested patients along observed and unobserved dimensions. We propose a Bayesian model class which captures this setting. The purpose of the model is to accurately estimate risk for both tested and untested patients. Estimating this model is challenging due to the wide range of possibilities for untested patients. To address this, we propose two domain constraints which are plausible in health settings: a prevalence constraint, where the overall disease prevalence is known, and an expertise constraint, where the human decision-maker deviates from purely risk-based decision-making only along a constrained feature set. We show theoretically and on synthetic data that domain constraints improve parameter inference. We apply our model to a case study of cancer risk prediction, showing that the model's inferred risk predicts cancer diagnoses, its inferred testing policy captures known public health policies, and it can identify suboptimalities in test allocation. Though our case study is in healthcare, our analysis reveals a general class of domain constraints which can improve model estimation in many settings.

Research in decoding visual information from the brain, particularly through the non-invasive fMRI method, is rapidly progressing. The challenge arises from the limited data availability and the low signal-to-noise ratio of fMRI signals, leading to a low-precision task of fMRI-to-image retrieval. State-of-the-art MindEye remarkably improves fMRI-to-image retrieval performance by leveraging a deep MLP with a high parameter count orders of magnitude, i.e., a 996M MLP Backbone per subject, to align fMRI embeddings to the final hidden layer of CLIP's vision transformer. However, significant individual variations exist among subjects, even within identical experimental setups, mandating the training of subject-specific models. The substantial parameters pose significant challenges in deploying fMRI decoding on practical devices, especially with the necessitating of specific models for each subject. To this end, we propose Lite-Mind, a lightweight, efficient, and versatile brain representation network based on discrete Fourier transform, that efficiently aligns fMRI voxels to fine-grained information of CLIP. Our experiments demonstrate that Lite-Mind achieves an impressive 94.3% fMRI-to-image retrieval accuracy on the NSD dataset for Subject 1, with 98.7% fewer parameters than MindEye. Lite-Mind is also proven to be able to be migrated to smaller brain datasets and establishes a new state-of-the-art for zero-shot classification on the GOD dataset. The code is available at https://github.com/gongzix/Lite-Mind.

Automated classification of animal sounds is a prerequisite for large-scale monitoring of biodiversity. Convolutional Neural Networks (CNNs) are among the most promising algorithms but they are slow, often achieve poor classification in the field and typically require large training data sets. Our objective was to design CNNs that are fast at inference time and achieve good classification performance while learning from moderate-sized data. Recordings from a rainforest ecosystem were used. Start and end-point of sounds from 20 bird species were manually annotated. Spectrograms from 10 second segments were used as CNN input. We designed simple CNNs with a frequency unwrapping layer (SIMP-FU models) such that any output unit was connected to all spectrogram frequencies but only to a sub-region of time, the Receptive Field (RF). Our models allowed experimentation with different RF durations. Models either used the time-indexed labels that encode start and end-point of sounds or simpler segment-level labels. Models learning from time-indexed labels performed considerably better than their segment-level counterparts. Best classification performances was achieved for models with intermediate RF duration of 1.5 seconds. The best SIMP-FU models achieved AUCs over 0.95 in 18 of 20 classes on the test set. On compact low-cost hardware the best SIMP-FU models evaluated up to seven times faster than real-time data acquisition. RF duration was a major driver of classification performance. The optimum of 1.5 s was in the same range as the duration of the sounds. Our models achieved good classification performance while learning from moderate-sized training data. This is explained by the usage of time-indexed labels during training and adequately sized RF. Results confirm the feasibility of deploying small CNNs with good classification performance on compact low-cost devices.

We present a framework for constraining the automatic sequential generation of equations to obey the rules of dimensional analysis by construction. Combining this approach with reinforcement learning, we built $\Phi$-SO, a Physical Symbolic Optimization method for recovering analytical functions from physical data leveraging units constraints. Our symbolic regression algorithm achieves state-of-the-art results in contexts in which variables and constants have known physical units, outperforming all other methods on SRBench's Feynman benchmark in the presence of noise (exceeding 0.1%) and showing resilience even in the presence of significant (10%) levels of noise.

Diffusion models have achieved state-of-the-art results on many modalities including images, speech, and video. However, existing models are not tailored to support remote sensing data, which is widely used in important applications including environmental monitoring and crop-yield prediction. Satellite images are significantly different from natural images - they can be multi-spectral, irregularly sampled across time - and existing diffusion models trained on images from the Web do not support them. Furthermore, remote sensing data is inherently spatio-temporal, requiring conditional generation tasks not supported by traditional methods based on captions or images. In this paper, we present DiffusionSat, to date the largest generative foundation model trained on a collection of publicly available large, high-resolution remote sensing datasets. As text-based captions are sparsely available for satellite images, we incorporate the associated metadata such as geolocation as conditioning information. Our method produces realistic samples and can be used to solve multiple generative tasks including temporal generation, superresolution given multi-spectral inputs and in-painting. Our method outperforms previous state-of-the-art methods for satellite image generation and is the first large-scale generative foundation model for satellite imagery. Diffusion models have achieved state of the art results in image generation (Sohl-Dickstein et al., 2015; Ho et al., 2020; Dhariwal & Nichol, 2021; Kingma et al., 2021; Song & Ermon, 2019; 2020). Large scale models such as Stable Diffusion Rombach et al. (2022) (SD) have been trained on Internet-scale image-text datasets to generate high-resolution images from user-provided captions.

Extractive question answering (QA) systems can enable physicians and researchers to query medical records, a foundational capability for designing clinical studies and understanding patient medical history. However, building these systems typically requires expert-annotated QA pairs. Large language models (LLMs), which can perform extractive QA, depend on high quality data in their prompts, specialized for the application domain. We introduce a novel approach, XAIQA, for generating synthetic QA pairs at scale from data naturally available in electronic health records. Our method uses the idea of a classification model explainer to generate questions and answers about medical concepts corresponding to medical codes. In an expert evaluation with two physicians, our method identifies $2.2\times$ more semantic matches and $3.8\times$ more clinical abbreviations than two popular approaches that use sentence transformers to create QA pairs. In an ML evaluation, adding our QA pairs improves performance of GPT-4 as an extractive QA model, including on difficult questions. In both the expert and ML evaluations, we examine trade-offs between our method and sentence transformers for QA pair generation depending on question difficulty.

Text-to-SQL simplifies database interactions by enabling non-experts to convert their natural language (NL) questions into Structured Query Language (SQL) queries. While recent advances in large language models (LLMs) have improved the zero-shot text-to-SQL paradigm, existing methods face scalability challenges when dealing with massive, dynamically changing databases. This paper introduces DBCopilot, a framework that addresses these challenges by employing a compact and flexible copilot model for routing across massive databases. Specifically, DBCopilot decouples the text-to-SQL process into schema routing and SQL generation, leveraging a lightweight sequence-to-sequence neural network-based router to formulate database connections and navigate natural language questions through databases and tables. The routed schemas and questions are then fed into LLMs for efficient SQL generation. Furthermore, DBCopilot also introduced a reverse schema-to-question generation paradigm, which can learn and adapt the router over massive databases automatically without requiring manual intervention. Experimental results demonstrate that DBCopilot is a scalable and effective solution for real-world text-to-SQL tasks, providing a significant advancement in handling large-scale schemas.

Being aware of important news is crucial for staying informed and making well-informed decisions efficiently. Natural Language Processing (NLP) approaches can significantly automate this process. This paper introduces the detection of important news, in a previously unexplored area, and presents a new benchmarking dataset (Khabarchin) for detecting important news in the Persian language. We define important news articles as those deemed significant for a considerable portion of society, capable of influencing their mindset or decision-making. The news articles are obtained from seven different prominent Persian news agencies, resulting in the annotation of 7,869 samples and the creation of the dataset. Two challenges of high disagreement and imbalance between classes were faced, and solutions were provided for them. We also propose several learning-based models, ranging from conventional machine learning to state-of-the-art transformer models, to tackle this task. Furthermore, we introduce the second task of important sentence detection in news articles, as they often come with a significant contextual length that makes it challenging for readers to identify important information. We identify these sentences in a weakly supervised manner.

In this work, we tackle the problem of unsupervised domain adaptation (UDA) for video action recognition. Our approach, which we call UNITE, uses an image teacher model to adapt a video student model to the target domain. UNITE first employs self-supervised pre-training to promote discriminative feature learning on target domain videos using a teacher-guided masked distillation objective. We then perform self-training on masked target data, using the video student model and image teacher model together to generate improved pseudolabels for unlabeled target videos. Our self-training process successfully leverages the strengths of both models to achieve strong transfer performance across domains. We evaluate our approach on multiple video domain adaptation benchmarks and observe significant improvements upon previously reported results.