Autonomous cars, such as those that Tesla Inc. builds, use complex deep neural networks to make instantaneous decisions based on real-time data from sensors on the vehicle. One major example of an autopilot mode feature is controlling the steering angle of the vehicle relative to the road. To build better self-driving models, we either need billions of miles of driving data or more complex models that train faster on lesser data. Conventional deep neural networks like the Convolutional Neural Network (CNN) are excellent at predicting steering angles by analyzing image data as input. This project entails constructing a hybrid neural network that combines classical and “quantum” layers in order to predict the steering angle for a car based on video data. Once a functional model is built, we will build a fully classical CNN and use it to compare against our hybrid CNN. The premise behind exploring this project lies in using quantum transfer learning methods similar to Variational Quantum Eigensolvers. We have a quantum circuit and an optimizer which optimizes parameters for a set of iterations till we find a desired value using quantum circuits. We will be using the layers in a deep neural network and transferring the weights as parameters to the quantum circuits. The quantum circuits will augment the classical neural network and help find the optimum steering angle based on images from a car's front camera. We shall build a deep CNN with several convolutional layers, and the following layers will be the variational quantum circuits. We will evaluate the performances based on the prediction of the steering angle and the time it takes for the neural networks to be trained.
