Resource-Efficient ML Models for Edge Computing

• Overview –

Heterogeneous resource-constrained devices vary in computational power, and each device possesses its own local data, which is often non-identically distributed. Traditional methods involve sending data to a central server for processing, raising privacy concerns. Federated learning addresses this by distributing surrogate models to clients, allowing training on local data, with only weight updates sent to the server for aggregation.

• Significance –

Despite good overall accuracy, global models struggle to classify minority classes effectively. Some approaches propose sharing side information between nodes to mitigate data heterogeneity, but this can lead to increased latency by waiting for stragglers.

• Contributions –

We  focus on improving the global model’s accuracy for minority classes by leveraging a detailed analysis of layer-wise relevance propagation. By identifying the dominant filters associated with specific minority classes, we introduce a scaling factor to mitigate the accuracy dip. Additionally, each client will adaptively determine its optimal quantization level based on various client parameters.

• Overview –

UAVs and UGVs primarily collect data in two formats: LiDAR and RGB. RGB data is semantically rich, providing detailed color and texture information, while LiDAR data is geometrically rich, capturing precise spatial and depth details. Relying on only one data type leads to the loss of critical complementary information from the other.

• Significance –

Since each device may capture a particular instance at different times, it is crucial to account for temporal information when fusing the various frames to ensure accurate feature fusion.

• Contributions –

Design an encoder-decoder architecture for devices-server respectively that enables dynamic spatio-temporal feature fusion.

• Overview –

Remote photoplethysmography (rPPG) is a non-contact method for estimating heart rate using regular video cameras. It leverages the periodic variation in skin color caused by blood volume changes in the skin with each heartbeat, which affects the light reflections captured by the camera. These subtle skin reflection (diffusion) changes are processed to extract heart rate information, offering a contactless, scalable alternative to traditional heart rate monitoring systems.

• Significance –

Regular heart rate monitoring is crucial for detecting early signs of cardiovascular diseases and other health conditions. rPPG holds immense potential for widespread use due to the ubiquity of camera sensors in smartphones, laptops, and other devices. Its contactless nature ensures user safety, making it ideal for healthcare environments where minimizing physical contact is essential, especially in infectious disease settings. By enabling regular and remote heart rate monitoring, rPPG can early detect cardiovascular issues and other health conditions, offering a convenient tool for patients and caregivers.

• Obstacles –

Despite its potential, rPPG faces several challenges. The diffusion signal captured in video data often has a low signal-to-noise ratio (SNR), making extracting accurate heart rate information difficult. External factors such as skin color, lighting conditions, camera sensor quality, and variations in camera-subject settings (e.g., distance, angle) add complexity and variability. Additionally, there are concerns about user privacy and data security, as rPPG systems rely on video footage of the individual, which could be misused if not properly safeguarded. Addressing these challenges requires developing robust algorithms capable of handling diverse real-world conditions while ensuring user privacy and safety.

• Overview –

Contactless respiratory rate (RR) estimation monitors subtle external movements of breathing-related organs such as the chest, shoulders, torso, and abdomen, caused by respiration. These movements are captured to estimate RR, providing continuous, non-invasive monitoring.

• Significance –

Continuous RR monitoring is critical for detecting conditions like sleep apnea and respiratory problems, especially in infants, older adults, and critically ill patients. This contactless approach enables regular, low-cost RR monitoring in a variety of settings, making it accessible and scalable for widespread use.

• Obstacles –

The subtle nature of respiration-induced movements results in low signal-to-noise ratio (SNR), complicating accurate detection. A multimodal approach is often necessary for effectiveness, and concerns about safety, privacy, and trustworthiness must be addressed to ensure secure and reliable systems.