In this project, we will build a neural network to evaluate the CIFAR-10 dataset using PyTorch. The model that is built in this notebook reached an accuracy of 72.52% on the test set of the CIFAR-10 dataset.
You are a new machine learning engineer at a self-driving car startup. Management is trying to decide whether to build or buy an object detection algorithm for objects that may be on the side of the road. They are considering buying a computer vision algorithm from a company called Detectocorp. Detectocorp’s algorithm claims a 70% accuracy rate on the CIFAR-10 dataset, a benchmark used to evaluate the state of the art for computer vision systems.
But before making this purchase, management wants to explore whether you can build an in-house solution that performs well. They have asked you to try your hand at creating a neural network that can classify arbitrary objects and potentially be fine-tuned on a larger dataset using transfer learning.
Our task is to build an image classifier using the CIFAR-10 dataset and evaluate its accuracy. Then we'll compare its performance to both Detectocorp’s algorithm (which achieved 70% accuracy) as well as the state of the art results detailed in the notebook—and make a recommendation to management about whether to build the solution in-house or buy the algorithm from Detectocorp.
The following are the main steps of this project.
- Explore and prepare the data for training and testing.
- Design and build your neural network.
- Train the neural network on the training set.
- Evaluate your network's performance on the test set.
- Make a recommendation on the build vs. buy business decision.