Multi-task learning
Example 1: Simplified autonomous driving example Self driving car need to detect:
- Pedestrians
- cars
- stop signs
- traffic lights ...
1 image can have multiple(4) labels, building a single NN, looking at each image and solving 4 problems at same time.
1 NN that resolve 4 problems performs better than 4 NN that solve 1 problem separately.
When multi-task learning makes sense
- Training on a set of tasks that could benefit from having shared lower-level features
- recognizing pedestrians, cars, stop signs
- Usually: Amount of data you have for each task is quite similar.
- Can train a big enough neural network to do well on all the tasks
In practice, multi-task learning is used much less often than transfer learning, if you have a small data set, use transfer learning, while if you have a big dataset, and each feature has similar amount of data, just use multi-task learning.