Designing neural networks under different user preferences
For a specific task, different users care about different domains, model sizes,
latency targets, accuracy levels, energy budgets, and deployment environments.
These requirements are often inconsistent, making manual neural network design difficult.
Task-specific
User-driven
Multi-objective
The challenge is not only achieving high accuracy, but also finding an architecture
that fits practical constraints such as speed, memory, power, and deployment platform.
Constraint signals merging into one architecture
Neural Network
Architecture
One design must satisfy multiple user-specific constraints at the same time.
Domain
Accuracy
Latency
Model Size
Energy Budget
Deployment