A5025842315- Autism Spectrum Disorder Research
- Child Development and Digital Technology
- Advanced Data Storage Technologies
- Emotion and Mood Recognition
- Infant Health and Development
- Parallel Computing and Optimization Techniques
- Human Pose and Action Recognition
- Privacy-Preserving Technologies in Data
- Ferroelectric and Negative Capacitance Devices
- Domain Adaptation and Few-Shot Learning
- Distributed systems and fault tolerance
- Assistive Technology in Communication and Mobility
- Mobile Crowdsensing and Crowdsourcing
- Age of Information Optimization
- Advanced Memory and Neural Computing
- Adversarial Robustness in Machine Learning
- 3D IC and TSV technologies
- Green IT and Sustainability
- Advanced Vision and Imaging
- Virology and Viral Diseases
- Reinforcement Learning in Robotics
- Low-power high-performance VLSI design
- Caching and Content Delivery
- Distributed and Parallel Computing Systems
- Image and Video Quality Assessment
ETH Zurich
2017-2024
Stanford University
2021-2023
Carnegie Mellon University
2011-2017
Machine Learning (ML) techniques have been rapidly adopted by smart Cyber-Physical Systems (CPS) and Internet-of-Things (IoT) due to their powerful decision-making capabilities. However, they are vulnerable various security reliability threats, at both hardware software levels, that compromise accuracy. These threats get aggravated in emerging edge ML devices stringent constraints terms of resources (e.g., compute, memory, power/energy), therefore cannot employ costly measures. Security,...
Activity recognition computer vision algorithms can be used to detect the presence of autism-related behaviors, including what are termed "restricted and repetitive behaviors", or stimming, by diagnostic instruments. Examples stimming include hand flapping, spinning, head banging. One most significant bottlenecks for implementing such classifiers is lack sufficiently large training sets human behavior specific pediatric developmental delays. The data that do exist usually recorded with a...
Background A formal autism diagnosis can be an inefficient and lengthy process. Families may wait several months or longer before receiving a for their child despite evidence that earlier intervention leads to better treatment outcomes. Digital technologies detect the presence of behaviors related scale access pediatric diagnoses. strong indicator is self-stimulatory such as hand flapping. Objective This study aims demonstrate feasibility deep learning detection flapping from unstructured...
Automated emotion classification could aid those who struggle to recognize emotions, including children with developmental behavioral conditions such as autism. However, most computer vision recognition models are trained on adult and therefore underperform when applied child faces.
Implementing automated facial expression recognition on mobile devices could provide an accessible diagnostic and therapeutic tool for those who struggle to recognize expressions, including children with developmental behavioral conditions such as autism. Despite recent advances in classifiers children, existing models are too computationally expensive smartphone use.
Processing-using-DRAM (PUD) is a paradigm where the analog operational properties of DRAM structures are used to perform bulk logic operations. While PUD promises high throughput at low energy and area cost, we uncover three limitations existing approaches that lead significant inefficiencies: (i) static data representation, i.e., 2's complement with fixed bit-precision, leading unnecessary computation over useless (i.e., inconsequential) data; (ii) support for only throughput-oriented...
RowHammer is a major read disturbance mechanism in DRAM where repeatedly accessing (hammering) row of cells (DRAM row) induces bitflips physically nearby rows (victim rows). To ensure robust operation, state-of-the-art mitigation mechanisms restore the charge potential victim (i.e., they perform preventive refresh or restoration). With newer chip generations, these more aggressively and cause larger performance, energy, area overheads. Therefore, it essential to develop better understanding...
Growing application memory demands and concurrent usage are making mobile device scarce. When pressure is high, current systems use a RAM-based compressed swap scheme (called ZRAM) to compress unused execution-related data anonymous in Linux) main memory. We observe that the state-of-the-art ZRAM prolongs relaunch latency wastes CPU time because it does not differentiate between hot cold or leverage different compression chunk sizes locality. make three new observations. 1) has levels of...
Hybrid storage systems (HSS) combine multiple devices with diverse characteristics to achieve high performance and capacity at low cost. The of an HSS highly depends on the effectiveness two key policies: (1) data-placement policy, which determines best-fit device for incoming data, (2) data-migration rearranges stored data across sustain performance. Prior works focus improving only placement or migration in HSS, leads sub-optimal Unfortunately, no prior work tries optimize both policies...
As the need for edge computing grows, many modern consumer devices now contain machine learning (ML) accelerators that can compute a wide range of neural network (NN) models while still fitting within tight resource constraints. We analyze commercial Edge TPU using 24 Google NN (including CNNs, LSTMs, transducers, and RCNNs), find accelerator suffers from three shortcomings, in terms computational throughput, energy efficiency, memory access handling. comprehensively study characteristics...
Digitally-delivered healthcare is well suited to address current inequities in the delivery of care due barriers access facilities. As COVID-19 pandemic phases out, we have a unique opportunity capitalize on familiarity with telemedicine approaches and continue advocate for mainstream adoption remote delivery. In this paper, specifically focus ability GuessWhat? smartphone-based charades-style gamified therapeutic intervention autism spectrum disorder (ASD) generate signal that distinguishes...
Modern DRAM-based systems suffer from significant energy and latency penalties due to conservative DRAM refresh standards. Volatile cells can retain information across a wide distribution of times ranging milliseconds many minutes, but each cell is currently refreshed every 64ms account for the extreme tail end retention time distribution, leading high overhead. Due poor technology scaling, this problem expected get worse in future device generations. Hence, current approach refreshing all...
Existing DRAM controllers employ rigid, nonadaptive scheduling and buffer management policies when servicing prefetch requests. Some treat prefetches the same as demand requests, others always prioritize demands over prefetches. However, none of these rigid result in best performance because they do not take into account usefulness If are useless, treating equally can lead to significant loss extra bandwidth consumption. In contrast, if useful, prioritizing hurt by reducing throughput...
Facial expression recognition (FER) is a critical computer vision task for variety of applications. Despite the widespread use FER, there dearth racially diverse facial emotion datasets which are enriched children, teens, and adults. To bridge this gap, we have built database using publicly available videos from TikTok, video-focused social networking service. We describe construction TikTok database. The dataset extracted 6428 scraped TikTok. consist 9392 distinct individuals labels 15...
Reinforcement Learning (RL) trains agents to learn optimal behavior by maximizing reward signals from experience datasets. However, RL training often faces memory limitations, leading execution latencies and prolonged times. To overcome this, SwiftRL explores Processing-In-Memory (PIM) architectures accelerate workloads. We achieve near-linear performance scaling implementing algorithms like Tabular Q-learning SARSA on UPMEM PIM systems optimizing for hardware. Our experiments OpenAI GYM...
Abstract Motivation Recent advances in sequencing technologies have stressed the critical role of sequence analysis algorithms and tools genomics healthcare research. In particular, alignment is a fundamental building block many pipelines frequently performance bottleneck both terms execution time memory usage. Classical are based on dynamic programming often require quadratic with respect to length. As result, classical fail scale increasing lengths quickly become memory-bound due...
We introduce Temporal consistency for Test-time adaptation (TempT), a novel method test-time on videos through the use of temporal coherence predictions across sequential frames as self-supervision signal. TempT is an approach with broad potential applications in computer vision tasks, including facial expression recognition (FER) videos. evaluate TempT's performance AffWild2 dataset. Our focuses solely unimodal visual aspect data and utilizes popular 2D CNN backbone, contrast to larger or...