Dynamic Compute Path Selection for Autonomous Vehicle Latency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in achieving faster reaction times due to the high compute cost and latency associated with their existing static deep learning compute paths, which can limit their ability to safely perform operations in real-time scenarios, especially when encountering events that require quick responses.
Innovation Solution
The system dynamically learns and selects compute paths within a neural network to determine the most suitable path for a particular scene or operation, intelligently activating and deactivating neurons and layers to reduce compute waste and latency, and implements an early exit strategy to reduce processing time and latency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If a static deep learning compute path is used, then the system structure is simple and reliable, but the reaction time is slow and compute resource usage is high
Solution Approach 1:
The patent implements dynamic compute path selection by transitioning from static to dynamic neural network architecture. The system dynamically determines which compute paths to activate based on real-time scene complexity and vehicle state, allowing the network to adapt its structure during operation. This enables faster reaction times by selecting simpler paths for straightforward scenarios while maintaining the capability for complex processing when needed.
Solution Approach 2:
The patent segments the neural network into multiple compute paths with different levels of complexity. Instead of a single monolithic processing path, the system divides the computation into separate pathways (e.g., first compute path, second compute path) that can be independently selected and activated. This segmentation allows the system to choose the most appropriate path length and complexity for each specific driving scenario, reducing unnecessary computation and improving reaction time.
2Productivity
If a static deep learning compute path is used, then the system is reliable and stable, but compute resource usage is high and latency is increased
Solution Approach 1:
The patent changes key parameters of the neural network dynamically, including the activation status of different compute paths, the number of active layers, and the complexity level of processing. By adjusting these parameters based on real-time scene analysis and vehicle operational state, the system optimizes compute resource allocation. This allows high productivity in simple scenarios while maintaining reliability through appropriate complexity adjustment in complex scenarios.
Solution Approach 2:
The patent applies partial action by activating only the necessary compute paths and neural network layers required for the current driving scenario. Instead of running the full computational graph, the system selectively executes partial paths (e.g., using only the first compute path for simple detections, or combining first and second compute paths for complex scenarios). This reduces unnecessary compute resource consumption while maintaining sufficient processing capability and stability.
3Loss of time
If the full neural network is executed, then processing accuracy is high, but processing time and latency are increased
Solution Approach 1:
The patent extracts and removes unnecessary computational steps from the full neural network execution. By analyzing the scene and determining the minimum required processing depth, the system extracts only the essential compute paths needed for accurate detection in the current context. For simple scenarios, this means taking out intermediate processing layers entirely, executing only the first compute path. For complex scenarios, it combines multiple extracted paths (first and second compute paths) to achieve sufficient accuracy without the full network overhead.
Data Source
AI summary
Systems and techniques are provided for learning deep learning compute paths. An example method can determine a neural network element of a first compute path within a neural network, the first compute path comprising neural network elements including the neural network element, the neural network element comprising a neural network layer and/or a neural network node; determine a second compute path comprising the neural network element and different neural network elements, the second compute path having a smaller size than the first compute path, a lower latency than the first compute path, and/or a smaller compute cost than the second compute path; determine, based on one or more conditions, whether to process data from the neural network element through the first compute path or the second compute path; and process the data from the neural network element through one of the first compute path or the second compute path.