Abstract
Despite surprising performance on zero-shot transfer, pre-training a large-scale multimodal model is often prohibitive as it requires a huge amount of data and computing resources. In this paper, we propose a method (BeamCLIP) that can effectively transfer the representations of a large pre-trained multimodal model (CLIP-ViT) into a small target model (e.g., ResNet-18). For unsupervised transfer, we introduce cross-modal similarity matching (CSM) that enables a student model to learn the representations of a teacher model by matching the relative similarity distribution across text prompt embeddings. To better encode the text prompts, we design context-based prompt augmentation (CPA) that can alleviate the lexical ambiguity of input text prompts. Our experiments show that unsupervised representation transfer of a pre-trained vision-language model enables a small ResNet-18 to achieve a better ImageNet-1K top-1 linear probe accuracy (66.2%) than vision-only self-supervised learning (SSL) methods (e.g., SimCLR: 51.8%, SwAV: 63.7%), while closing the gap with supervised learning (69.8%).
REaaS: Enabling Adversarially Robust Downstream Classifiers via Robust Encoder as a Service
Authors: Wenjie Qu, Jinyuan Jia, Neil Zhenqiang Gong
Subjects: Cryptography and Security (cs.CR); Machine Learning (cs.LG)
Abstract
Encoder as a service is an emerging cloud service. Specifically, a service provider first pre-trains an encoder (i.e., a general-purpose feature extractor) via either supervised learning or self-supervised learning and then deploys it as a cloud service API. A client queries the cloud service API to obtain feature vectors for its training/testing inputs when training/testing its classifier (called downstream classifier). A downstream classifier is vulnerable to adversarial examples, which are testing inputs with carefully crafted perturbation that the downstream classifier misclassifies. Therefore, in safety and security critical applications, a client aims to build a robust downstream classifier and certify its robustness guarantees against adversarial examples. What APIs should the cloud service provide, such that a client can use any certification method to certify the robustness of its downstream classifier against adversarial examples while minimizing the number of queries to the APIs? How can a service provider pre-train an encoder such that clients can build more certifiably robust downstream classifiers? We aim to answer the two questions in this work. For the first question, we show that the cloud service only needs to provide two APIs, which we carefully design, to enable a client to certify the robustness of its downstream classifier with a minimal number of queries to the APIs. For the second question, we show that an encoder pre-trained using a spectral-norm regularization term enables clients to build more robust downstream classifiers.
Learning the Relation between Similarity Loss and Clustering Loss in Self-Supervised Learning
Authors: Jidong Ge, Yuxiang Liu, Jie Gui, Lanting Fang, Ming Lin, James Tin-Yau Kwok, LiGuo Huang, Bin Luo
Subjects: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
Abstract
Self-supervised learning enables networks to learn discriminative features from massive data itself. Most state-of-the-art methods maximize the similarity between two augmentations of one image based on contrastive learning. By utilizing the consistency of two augmentations, the burden of manual annotations can be freed. Contrastive learning exploits instance-level information to learn robust features. However, the learned information is probably confined to different views of the same instance. In this paper, we attempt to leverage the similarity between two distinct images to boost representation in self-supervised learning. In contrast to instance-level information, the similarity between two distinct images may provide more useful information. Besides, we analyze the relation between similarity loss and feature-level cross-entropy loss. These two losses are essential for most deep learning methods. However, the relation between these two losses is not clear. Similarity loss helps obtain instance-level representation, while feature-level cross-entropy loss helps mine the similarity between two distinct images. We provide theoretical analyses and experiments to show that a suitable combination of these two losses can get state-of-the-art results.
Seamless Multimodal Biometrics for Continuous Personalised Wellbeing Monitoring
Authors: João Ribeiro Pinto
Subjects: Computer Vision and Pattern Recognition (cs.CV)
Abstract
Artificially intelligent perception is increasingly present in the lives of every one of us. Vehicles are no exception, (...) In the near future, pattern recognition will have an even stronger role in vehicles, as self-driving cars will require automated ways to understand what is happening around (and within) them and act accordingly. (...) This doctoral work focused on advancing in-vehicle sensing through the research of novel computer vision and pattern recognition methodologies for both biometrics and wellbeing monitoring. The main focus has been on electrocardiogram (ECG) biometrics, a trait well-known for its potential for seamless driver monitoring. Major efforts were devoted to achieving improved performance in identification and identity verification in off-the-person scenarios, well-known for increased noise and variability. Here, end-to-end deep learning ECG biometric solutions were proposed and important topics were addressed such as cross-database and long-term performance, waveform relevance through explainability, and interlead conversion. Face biometrics, a natural complement to the ECG in seamless unconstrained scenarios, was also studied in this work. The open challenges of masked face recognition and interpretability in biometrics were tackled in an effort to evolve towards algorithms that are more transparent, trustworthy, and robust to significant occlusions. Within the topic of wellbeing monitoring, improved solutions to multimodal emotion recognition in groups of people and activity/violence recognition in in-vehicle scenarios were proposed. At last, we also proposed a novel way to learn template security within end-to-end models, dismissing additional separate encryption processes, and a self-supervised learning approach tailored to sequential data, in order to ensure data security and optimal performance. (...)
Self-Supervised Time-to-Event Modeling with Structured Medical Records
Authors: Ethan Steinberg, Yizhe Xu, Jason Fries, Nigam Shah
Abstract
Time-to-event models (also known as survival models) are used in medicine and other fields for estimating the probability distribution of the time until a particular event occurs. While providing many advantages over traditional classification models, such as naturally handling censoring, time-to-event models require more parameters and are challenging to learn in settings with limited labeled training data. High censoring rates, common in events with long time horizons, further limit available training data and exacerbate the risk of overfitting. Existing methods, such as proportional hazard or accelerated failure time-based approaches, employ distributional assumptions to reduce parameter size, but they are vulnerable to model misspecification. In this work, we address these challenges with MOTOR, a self-supervised model that leverages temporal structure found in large-scale collections of timestamped, but largely unlabeled events, typical of electronic health record data. MOTOR defines a time-to-event pretraining task that naturally captures the probability distribution of event times, making it well-suited to applications in medicine. After pretraining on 8,192 tasks auto-generated from 2.7M patients (2.4B clinical events), we evaluate the performance of our pretrained model after fine-tuning to unseen time-to-event tasks. MOTOR-derived models improve upon current state-of-the-art C statistic performance by 6.6% and decrease training time (in wall time) by up to 8.2 times. We further improve sample efficiency, with adapted models matching current state-of-the-art performance using 95% less training data.
Data Discovery using Natural Language Questions via a Self-Supervised Approach
Abstract
Data discovery systems help users identify relevant data among large table collections. Users express their discovery needs with a program or a set of keywords. Users may express complex queries using programs but it requires expertise. Keyword search is accessible to a larger audience but limits the types of queries supported. An interesting approach is learned discovery systems which find tables given natural language questions. Unfortunately, these systems require a training dataset for each table collection. And because collecting training data is expensive, this limits their adoption. In this paper, we introduce a self-supervised approach to assemble training datasets and train learned discovery systems without human intervention. It requires addressing several challenges, including the design of self-supervised strategies for data discovery, table representation strategies to feed to the models, and relevance models that work well with the synthetically generated questions. We combine all the above contributions into a system, S2LD, that solves the problem end to end. The evaluation results demonstrate the new techniques outperform state-of-the-art approaches on wellknown benchmarks. All in all, the technique is a stepping stone towards building learned discovery systems. The code is open-sourced at https://github.com/TheDataStation/open_table_discovery.
Keyword: vision transformer
STPrivacy: Spatio-Temporal Tubelet Sparsification and Anonymization for Privacy-preserving Action Recognition
Authors: Ming Li, Jun Liu, Hehe Fan, Jia-Wei Liu, Jiahe Li, Mike Zheng Shou, Jussi Keppo
Subjects: Computer Vision and Pattern Recognition (cs.CV)
Abstract
Recently privacy-preserving action recognition (PPAR) has been becoming an appealing video understanding problem. Nevertheless, existing works focus on the frame-level (spatial) privacy preservation, ignoring the privacy leakage from a whole video and destroying the temporal continuity of actions. In this paper, we present a novel PPAR paradigm, i.e., performing privacy preservation from both spatial and temporal perspectives, and propose a STPrivacy framework. For the first time, our STPrivacy applies vision Transformers to PPAR and regards a video as a sequence of spatio-temporal tubelets, showing outstanding advantages over previous convolutional methods. Specifically, our STPrivacy adaptively treats privacy-containing tubelets in two different manners. The tubelets irrelevant to actions are directly abandoned, i.e., sparsification, and not published for subsequent tasks. In contrast, those highly involved in actions are anonymized, i.e., anonymization, to remove private information. These two transformation mechanisms are complementary and simultaneously optimized in our unified framework. Because there is no large-scale benchmarks, we annotate five privacy attributes for two of the most popular action recognition datasets, i.e., HMDB51 and UCF101, and conduct extensive experiments on them. Moreover, to verify the generalization ability of our STPrivacy, we further introduce a privacy-preserving facial expression recognition task and conduct experiments on a large-scale video facial attributes dataset, i.e., Celeb-VHQ. The thorough comparisons and visualization analysis demonstrate our significant superiority over existing works. The appendix contains more details and visualizations.
Keyword: multimodal
Multimodal Lyrics-Rhythm Matching
Authors: Callie C. Liao, Duoduo Liao, Jesse Guessford
Abstract
Despite the recent increase in research on artificial intelligence for music, prominent correlations between key components of lyrics and rhythm such as keywords, stressed syllables, and strong beats are not frequently studied. Ths is likely due to challenges such as audio misalignment, inaccuracies in syllabic identification, and most importantly, the need for cross-disciplinary knowledge. To address this lack of research, we propose a novel multimodal lyrics-rhythm matching approach in this paper that specifically matches key components of lyrics and music with each other without any language limitations. We use audio instead of sheet music with readily available metadata, which creates more challenges yet increases the application flexibility of our method. Furthermore, our approach creatively generates several patterns involving various multimodalities, including music strong beats, lyrical syllables, auditory changes in a singer's pronunciation, and especially lyrical keywords, which are utilized for matching key lyrical elements with key rhythmic elements. This advantageous approach not only provides a unique way to study auditory lyrics-rhythm correlations including efficient rhythm-based audio alignment algorithms, but also bridges computational linguistics with music as well as music cognition. Our experimental results reveal an 0.81 probability of matching on average, and around 30% of the songs have a probability of 0.9 or higher of keywords landing on strong beats, including 12% of the songs with a perfect landing. Also, the similarity metrics are used to evaluate the correlation between lyrics and rhythm. It shows that nearly 50% of the songs have 0.70 similarity or higher. In conclusion, our approach contributes significantly to the lyrics-rhythm relationship by computationally unveiling insightful correlations.
Transferring Pre-trained Multimodal Representations with Cross-modal Similarity Matching
Authors: Byoungjip Kim, Sungik Choi, Dasol Hwang, Moontae Lee, Honglak Lee
Subjects: Machine Learning (cs.LG); Computer Vision and Pattern Recognition (cs.CV)
Abstract
Despite surprising performance on zero-shot transfer, pre-training a large-scale multimodal model is often prohibitive as it requires a huge amount of data and computing resources. In this paper, we propose a method (BeamCLIP) that can effectively transfer the representations of a large pre-trained multimodal model (CLIP-ViT) into a small target model (e.g., ResNet-18). For unsupervised transfer, we introduce cross-modal similarity matching (CSM) that enables a student model to learn the representations of a teacher model by matching the relative similarity distribution across text prompt embeddings. To better encode the text prompts, we design context-based prompt augmentation (CPA) that can alleviate the lexical ambiguity of input text prompts. Our experiments show that unsupervised representation transfer of a pre-trained vision-language model enables a small ResNet-18 to achieve a better ImageNet-1K top-1 linear probe accuracy (66.2%) than vision-only self-supervised learning (SSL) methods (e.g., SimCLR: 51.8%, SwAV: 63.7%), while closing the gap with supervised learning (69.8%).
Towards early prediction of neurodevelopmental disorders: Computational model for Face Touch and Self-adaptors in Infants
Authors: Bruno Tafur, Marwa Mahmoud, Staci Weiss
Subjects: Computer Vision and Pattern Recognition (cs.CV)
Abstract
Infants' neurological development is heavily influenced by their motor skills. Evaluating a baby's movements is key to understanding possible risks of developmental disorders in their growth. Previous research in psychology has shown that measuring specific movements or gestures such as face touches in babies is essential to analyse how babies understand themselves and their context. This research proposes the first automatic approach that detects face touches from video recordings by tracking infants' movements and gestures. The study uses a multimodal feature fusion approach mixing spatial and temporal features and exploits skeleton tracking information to generate more than 170 aggregated features of hand, face and body. This research proposes data-driven machine learning models for the detection and classification of face touch in infants. We used cross dataset testing to evaluate our proposed models. The models achieved 87.0% accuracy in detecting face touches and 71.4% macro-average accuracy in detecting specific face touch locations with significant improvements over Zero Rule and uniform random chance baselines. Moreover, we show that when we run our model to extract face touch frequencies of a larger dataset, we can predict the development of fine motor skills during the first 5 months after birth.
Seamless Multimodal Biometrics for Continuous Personalised Wellbeing Monitoring
Authors: João Ribeiro Pinto
Subjects: Computer Vision and Pattern Recognition (cs.CV)
Abstract
Artificially intelligent perception is increasingly present in the lives of every one of us. Vehicles are no exception, (...) In the near future, pattern recognition will have an even stronger role in vehicles, as self-driving cars will require automated ways to understand what is happening around (and within) them and act accordingly. (...) This doctoral work focused on advancing in-vehicle sensing through the research of novel computer vision and pattern recognition methodologies for both biometrics and wellbeing monitoring. The main focus has been on electrocardiogram (ECG) biometrics, a trait well-known for its potential for seamless driver monitoring. Major efforts were devoted to achieving improved performance in identification and identity verification in off-the-person scenarios, well-known for increased noise and variability. Here, end-to-end deep learning ECG biometric solutions were proposed and important topics were addressed such as cross-database and long-term performance, waveform relevance through explainability, and interlead conversion. Face biometrics, a natural complement to the ECG in seamless unconstrained scenarios, was also studied in this work. The open challenges of masked face recognition and interpretability in biometrics were tackled in an effort to evolve towards algorithms that are more transparent, trustworthy, and robust to significant occlusions. Within the topic of wellbeing monitoring, improved solutions to multimodal emotion recognition in groups of people and activity/violence recognition in in-vehicle scenarios were proposed. At last, we also proposed a novel way to learn template security within end-to-end models, dismissing additional separate encryption processes, and a self-supervised learning approach tailored to sequential data, in order to ensure data security and optimal performance. (...)
Abstract
Multimodal learning can benefit from the representation power of pretrained Large Language Models (LLMs). However, state-of-the-art transformer based LLMs often ignore negations in natural language and there is no existing benchmark to quantitatively evaluate whether multimodal transformers inherit this weakness. In this study, we present a new multimodal question answering (QA) benchmark adapted from labeled music videos in AudioSet (Gemmeke et al., 2017) with the goal of systematically evaluating if multimodal transformers can perform complex reasoning to recognize new concepts as negation of previously learned concepts. We show that with standard fine-tuning approach multimodal transformers are still incapable of correctly interpreting negation irrespective of model size. However, our experiments demonstrate that augmenting the original training task distributions with negated QA examples allow the model to reliably reason with negation. To do this, we describe a novel data generation procedure that prompts the 540B-parameter PaLM model to automatically generate negated QA examples as compositions of easily accessible video tags. The generated examples contain more natural linguistic patterns and the gains compared to template-based task augmentation approach are significant.
Universal Multimodal Representation for Language Understanding
Abstract
Representation learning is the foundation of natural language processing (NLP). This work presents new methods to employ visual information as assistant signals to general NLP tasks. For each sentence, we first retrieve a flexible number of images either from a light topic-image lookup table extracted over the existing sentence-image pairs or a shared cross-modal embedding space that is pre-trained on out-of-shelf text-image pairs. Then, the text and images are encoded by a Transformer encoder and convolutional neural network, respectively. The two sequences of representations are further fused by an attention layer for the interaction of the two modalities. In this study, the retrieval process is controllable and flexible. The universal visual representation overcomes the lack of large-scale bilingual sentence-image pairs. Our method can be easily applied to text-only tasks without manually annotated multimodal parallel corpora. We apply the proposed method to a wide range of natural language generation and understanding tasks, including neural machine translation, natural language inference, and semantic similarity. Experimental results show that our method is generally effective for different tasks and languages. Analysis indicates that the visual signals enrich textual representations of content words, provide fine-grained grounding information about the relationship between concepts and events, and potentially conduce to disambiguation.
Learning Bidirectional Action-Language Translation with Limited Supervision and Incongruent Extra Input
Authors: Ozan Özdemir, Matthias Kerzel, Cornelius Weber, Jae Hee Lee, Muhammad Burhan Hafez, Patrick Bruns, Stefan Wermter
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Neural and Evolutionary Computing (cs.NE); Robotics (cs.RO)
Abstract
Human infant learning happens during exploration of the environment, by interaction with objects, and by listening to and repeating utterances casually, which is analogous to unsupervised learning. Only occasionally, a learning infant would receive a matching verbal description of an action it is committing, which is similar to supervised learning. Such a learning mechanism can be mimicked with deep learning. We model this weakly supervised learning paradigm using our Paired Gated Autoencoders (PGAE) model, which combines an action and a language autoencoder. After observing a performance drop when reducing the proportion of supervised training, we introduce the Paired Transformed Autoencoders (PTAE) model, using Transformer-based crossmodal attention. PTAE achieves significantly higher accuracy in language-to-action and action-to-language translations, particularly in realistic but difficult cases when only few supervised training samples are available. We also test whether the trained model behaves realistically with conflicting multimodal input. In accordance with the concept of incongruence in psychology, conflict deteriorates the model output. Conflicting action input has a more severe impact than conflicting language input, and more conflicting features lead to larger interference. PTAE can be trained on mostly unlabelled data where labeled data is scarce, and it behaves plausibly when tested with incongruent input.
Keyword: CLIP
Transferring Pre-trained Multimodal Representations with Cross-modal Similarity Matching
Authors: Byoungjip Kim, Sungik Choi, Dasol Hwang, Moontae Lee, Honglak Lee
Subjects: Machine Learning (cs.LG); Computer Vision and Pattern Recognition (cs.CV)
Abstract
Despite surprising performance on zero-shot transfer, pre-training a large-scale multimodal model is often prohibitive as it requires a huge amount of data and computing resources. In this paper, we propose a method (BeamCLIP) that can effectively transfer the representations of a large pre-trained multimodal model (CLIP-ViT) into a small target model (e.g., ResNet-18). For unsupervised transfer, we introduce cross-modal similarity matching (CSM) that enables a student model to learn the representations of a teacher model by matching the relative similarity distribution across text prompt embeddings. To better encode the text prompts, we design context-based prompt augmentation (CPA) that can alleviate the lexical ambiguity of input text prompts. Our experiments show that unsupervised representation transfer of a pre-trained vision-language model enables a small ResNet-18 to achieve a better ImageNet-1K top-1 linear probe accuracy (66.2%) than vision-only self-supervised learning (SSL) methods (e.g., SimCLR: 51.8%, SwAV: 63.7%), while closing the gap with supervised learning (69.8%).
Keyword: metric learning
There is no result
Keyword: image retrieval
There is no result
Keyword: self-supervised
Transferring Pre-trained Multimodal Representations with Cross-modal Similarity Matching
REaaS: Enabling Adversarially Robust Downstream Classifiers via Robust Encoder as a Service
Learning the Relation between Similarity Loss and Clustering Loss in Self-Supervised Learning
Seamless Multimodal Biometrics for Continuous Personalised Wellbeing Monitoring
Self-Supervised Time-to-Event Modeling with Structured Medical Records
Data Discovery using Natural Language Questions via a Self-Supervised Approach
Keyword: vision transformer
STPrivacy: Spatio-Temporal Tubelet Sparsification and Anonymization for Privacy-preserving Action Recognition
Keyword: multimodal
Multimodal Lyrics-Rhythm Matching
Transferring Pre-trained Multimodal Representations with Cross-modal Similarity Matching
Towards early prediction of neurodevelopmental disorders: Computational model for Face Touch and Self-adaptors in Infants
Seamless Multimodal Biometrics for Continuous Personalised Wellbeing Monitoring
MAQA: A Multimodal QA Benchmark for Negation
Universal Multimodal Representation for Language Understanding
Learning Bidirectional Action-Language Translation with Limited Supervision and Incongruent Extra Input
Keyword: CLIP
Transferring Pre-trained Multimodal Representations with Cross-modal Similarity Matching
Keyword: DALLE
There is no result