Recent improvements in deep learning and hardware support offer a new breakthrough mobile sensing; we could enjoy context-aware services healthcare on device powered by artificial intelligence. However, most related studies perform well only with certain level of similarity between trained target data distribution, while practice, specific user's behaviors make sensor inputs different. Consequently, the performance such applications might suffer diverse user conditions as training models...

While smartphones have enriched our lives with diverse applications and functionalities, the user experience still often involves manual cumbersome inputs. To purchase a bottle of water for instance, must locate an e-commerce app, type keyword search, select right item from list, finally place order. This process could be greatly simplified if smartphone identifies object interest automatically executes preferred actions object. We present Knocker that when simply knocks on smartphone. The...

Various automated eating detection wearables have been proposed to monitor food intakes. While these systems overcome the forgetfulness of manual user journaling, they typically show low accuracy at outside-the-lab environments or intrusive form-factors (e.g., headgear). Eyeglasses are emerging as a socially-acceptable wearable, but existing approaches require custom-built frames and consume large power. We propose MyDJ, an system that could be attached any eyeglass frame. MyDJ achieves...

Test-time adaptation (TTA) is an emerging paradigm that addresses distributional shifts between training and testing phases without additional data acquisition or labeling cost; only unlabeled test streams are used for continual model adaptation. Previous TTA schemes assume the samples independent identically distributed (i.i.d.), even though they often temporally correlated (non-i.i.d.) in application scenarios, e.g., autonomous driving. We discover most existing methods fail dramatically...

As emojis are increasingly used in everyday online communication such as messaging, email, and social networks, various techniques have attempted to improve the user experience communicating emotions information through emojis. Emoji recommendation is one example which machine learning applied predict about select, based on user’s current input message. Although emoji suggestion helps users identify select right among a plethora of emojis, analyzing only single sentence for this purpose has...

Many applications utilize sensors in mobile devices and machine learning to provide novel services. However, various factors such as different users, devices, environments impact the performance of applications, thus making domain shift (i.e., distributional between training target domain) a critical issue sensing. Despite attempts adaptation solve this challenging problem, their is unreliable due complex interplay among diverse factors. In principle, uncertainty can be identified redeemed...

Speech emotion recognition (SER) models typically rely on costly human-labeled data for training, making scaling methods to large speech datasets and nuanced taxonomies difficult. We present LanSER, a method that enables the use of unlabeled by inferring weak labels via pre-trained language through weakly-supervised learning. For constrained taxonomy, we textual entailment approach selects an label with highest score transcript extracted automatic recognition. Our experimental results show...

Many applications utilize sensors on mobile devices and apply deep learning for diverse applications. However, they have rarely enjoyed mainstream adoption due to many different <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">individual conditions</i> users encounter. Individual conditions are characterized by users' unique behaviors carry, which collectively make sensor inputs different. It is impractical train countless individual...

While people primarily communicate with text in mobile chat applications, they are increasingly using visual elements such as images, emojis, and memes. Using could help users clearly make chatting experience enjoyable. However, finding inserting contextually appropriate images during the can be both tedious distracting. We introduce MilliCat, a real-time image suggestion system that recommends match content within application (i.e., autocomplete images). MilliCat combines natural language...

The advent of tiny AI accelerators opens opportunities for deep neural network deployment at the extreme edge, offering reduced latency, lower power cost, and improved privacy in on-device ML inference. Despite these advancements, challenges persist due to inherent limitations accelerators, such as restricted onboard memory single-device focus. This paper introduces Synergy, a system that dynamically composes multi-tenant models, effectively addressing tinyML's critical increasing demand AI....

Microcontrollers (MCUs) have been deployed on numerous IoT devices due to their compact sizes and low costs. MCUs are capable of capturing sensor data processing them. However, computational power, applications with deep neural networks (DNNs) limited. In this paper, we propose MiCrowd, a floating population measurement system tiny DNNs running since the essential value in urban planning business. Moreover, MiCrowd addresses following important challenges: (1) privacy issues, (2)...

We argue for research on identifying opportune moments remote computer-mediated interactions with home-alone dogs. analyze the behavior of pet dogs to find specific situations where positive interaction between dog and toys is more likely when might induce stress. highlight importance considering timing potential benefits it brings effectiveness interaction, leading greater satisfaction engagement both owner.

We present MIRROR, an on-device video virtual try-on (VTO) system that provides realistic, private, and rapid experiences in mobile clothes shopping. Despite recent advancements generative adversarial networks (GANs) for VTO, designing MIRROR involves two challenges: (1) data discrepancy due to restricted training miss various poses, body sizes, backgrounds (2) local computation overhead uses up 24% of battery converting only a single video. To alleviate the problems, we propose...

Deep learning has enabled personal and IoT devices to rethink microphones as a multi-purpose sensor for understanding conversation the surrounding environment. This resulted in proliferation of Voice Controllable Systems (VCS) around us. The increasing popularity such systems is also prone attracting miscreants, who often want take advantage VCS without knowledge user. Consequently, robustness VCS, especially under adversarial attacks, become an important research topic. Although there...

We use smartphones and their apps for almost every daily activity. For instance, to purchase a bottle of water online, user has unlock the smartphone, find right e-commerce app, search name product, finally place an order. This procedure requires manual, often cumbersome, input user, but could be significantly simplified if smartphone can identify object automatically process this routine. present Knocker, identification technique that only uses commercial off-the-shelf smartphones. The...

We present Reeboc that combines machine learning and k-means clustering to analyze the conversation of a chat, extract different emotions or topics conversation, recommend emojis represent various contexts user. Instead simply analyzing single input sentence, we consider recent sentences exchanged in conversation. performed user study with 17 participants 8 groups realistic mobile chat environment. Participants spent least amount time identifying selecting their choice (38% faster than...

Test-time adaptation (TTA) aims to address distributional shifts between training and testing data using only unlabeled test streams for continual model adaptation. However, most TTA methods assume benign streams, while samples could be unexpectedly diverse in the wild. For instance, an unseen object or noise appear autonomous driving. This leads a new threat existing algorithms; we found that prior algorithms suffer from those noisy as they blindly adapt incoming samples. To this problem,...

The miniaturization of AI accelerators is paving the way for next-generation wearable applications within technologies. We introduce Mojito, an AI-native runtime with advanced MLOps designed to facilitate development and deployment these on devices. It emphasizes necessity dynamic orchestration distributed resources equipped ultra-low-power overcome challenges associated unpredictable environments. Through its innovative approaches, Mojito demonstrates how future technologies can evolve be...

Test-time adaptation (TTA) has emerged as a viable solution to adapt pre-trained models domain shifts using unlabeled test data. However, TTA faces challenges of failures due its reliance on blind unknown samples in dynamic scenarios. Traditional methods for out-of-distribution performance estimation are limited by unrealistic assumptions the context, such requiring labeled data or re-training models. To address this issue, we propose AETTA, label-free accuracy algorithm TTA. We prediction...