Cloud with Arrows IconTheme 3: Distributed Intelligence


Performing coordinated cognitive tasks on data collected from distributed sensors is a key challenge in many applications (e.g., a swarm of drones collectively tracking an object). With the widespread proliferation of Internet-of-Things (IoT) devices, the volume, velocity, and variety of data that they generate will rapidly increase.


 

State-of-the-Art:
Cloud-enabled Intelligence

  • Centralized training in cloud
  • Inference entirely in cloud or entirely on edge device
  • Cognition requires uncompressed data
  • Algorithms agnostic to distributed context require high communication

 

Future Distributed Intelligence

  • Distributed training and inference in hierarchical (edge-hub-cloud) or peer-to-peer setting
  • Cognition on compressed data
  • Cognition on incomplete and unreliable data
  • Context-aware distributed cognition reduces communication

Fig. 3.1: Key research directions in Distributed Intelligence to be explored in Theme 3


The value of the IoT lies in making sense of these exploding streams of data and turning them into actionable insights. However, the current approach of uploading all the data to the cloud for analysis will become infeasible due to the high communication energy required, privacy concerns, and bandwidth and latency constraints when mission-critical real-time response is required. The use of in-sensor and near-sensor data processing and analytics is expected to play a crucial role. Nevertheless, performing learning and inference entirely at the edge may also not be feasible due to the computational and energy limitations of most edge devices. In some scenarios, we will develop autonomous intelligent systems to form dynamic peer-to-peer networks that lack a centralized and coordinated infrastructure, with significant network and node unreliability. These attributes lead to significant challenges in learning and inference. We will develop break-through technologies to enable distributed intelligence with emphasis on energy-efficiency, adaptability and performance. We will pursue the following three tasks: 3.1: Distributed learning and inference; 3.2: Cognition on compressed and unreliable data; 3.3: Context-aware distributed cognition.

Theme Leader

Principal Investigators (PIs)

Cloud with Arrows IconTheme 3: Distributed Intelligence


Performing coordinated cognitive tasks on data collected from distributed sensors is a key challenge in many applications (e.g., a swarm of drones collectively tracking an object). With the widespread proliferation of Internet-of-Things (IoT) devices, the volume, velocity, and variety of data that they generate will rapidly increase.


 

State-of-the-Art:
Cloud-enabled Intelligence

  • Centralized training in cloud
  • Inference entirely in cloud or entirely on edge device
  • Cognition requires uncompressed data
  • Algorithms agnostic to distributed context require high communication

 

Future Distributed Intelligence

  • Distributed training and inference in hierarchical (edge-hub-cloud) or peer-to-peer setting
  • Cognition on compressed data
  • Cognition on incomplete and unreliable data
  • Context-aware distributed cognition reduces communication

Fig. 3.1: Key research directions in Distributed Intelligence to be explored in Theme 3


The value of the IoT lies in making sense of these exploding streams of data and turning them into actionable insights. However, the current approach of uploading all the data to the cloud for analysis will become infeasible due to the high communication energy required, privacy concerns, and bandwidth and latency constraints when mission-critical real-time response is required. The use of in-sensor and near-sensor data processing and analytics is expected to play a crucial role. Nevertheless, performing learning and inference entirely at the edge may also not be feasible due to the computational and energy limitations of most edge devices. In some scenarios, we will develop autonomous intelligent systems to form dynamic peer-to-peer networks that lack a centralized and coordinated infrastructure, with significant network and node unreliability. These attributes lead to significant challenges in learning and inference. We will develop break-through technologies to enable distributed intelligence with emphasis on energy-efficiency, adaptability and performance. We will pursue the following three tasks: 3.1: Distributed learning and inference; 3.2: Cognition on compressed and unreliable data; 3.3: Context-aware distributed cognition.

Theme Leader

Principal Investigators (PIs)