Choosing a design for an indoor delivery robot with little data.
Case context: You are leading a team that is building an indoor delivery robot for a hospital corridor system. You only have 280 labeled runs from a prototype robot, so the dataset is small. You are deciding between two machine learning designs: a single end-to-end model that maps sensor readings directly to motion commands, or a pipeline that first detects people, then detects obstacles such as carts and doors, and then plans a safe route.
Question: Using the idea that some functions are easier to learn than others, which design should you choose if you need the system to work well with limited data? Explain your choice.
Sample answer: You should choose the pipeline design. It breaks the task into simpler subtasks: detecting people, detecting obstacles, and planning a route. Each subtask is easier to learn than a single model that goes from raw sensor input straight to motion commands. Because simpler functions usually need less training data, the pipeline is the better choice when data is scarce.
Key points:
- Choose the pipeline design instead of the end-to-end design.
- The pipeline contains three simpler steps: detect people, detect obstacles, and plan a route.
- Simpler steps are easier to learn than one very complex mapping.
- Easier-to-learn steps need less data, which matters when the dataset is small.
Rubric: The answer must choose the pipeline design and justify it by explaining that the three subtasks are simpler functions that can be learned with less data than a direct end-to-end mapping.
0
1
Tags
Machine Learning
Deep Learning
Supervised Learning
Dive into Deep Learning @ D2L
Data Science
Machine Learning Strategy
Machine Learning Yearning @ DeepLearning.AI
Related
Main Modules in a Modular Self-Driving System
True or False: In a self-driving system broken into stages, each stage is usually a simpler learning problem and can often be trained with less data than a single end-to-end model.
A warehouse robot workflow is divided into four main stages: scan the shelf, choose a path, travel to the target, and confirm the pickup. How many main stages are there?
Which three subtasks are the core stages of a self-driving navigation pipeline?
A modular speech-recognition pipeline usually needs more labeled data for each subtask than a single end-to-end model would need.
A delivery-drone pipeline breaks the task into _____ main steps.
Match each step in a warehouse robot pipeline to its main purpose.
Put the driving pipeline steps in the order of a simple perception-to-planning system.
Why does a modular self-driving system often need less labeled data for each module than a single end-to-end model?
A self-driving pipeline that separates perception, planning, and control gives the learning system an explicit breakdown of the driving task.
The third step in the drone navigation pipeline, after spotting buildings and other drones, is to _____ for the drone.
Match each concept to its meaning in a warehouse robot navigation pipeline.
Order the reasoning steps a developer follows when deciding whether to use a perception-and-planning pipeline for a self-driving car, from first to last.
Why breaking a robot navigation task into subtasks can reduce the amount of training data needed.
Choosing a design for an indoor delivery robot with little data.
Relative learning difficulty of modular driving tasks versus end-to-end control