Nice project. May I ask what are the major differences between this project and the existing autopilot systems in term of the architecture of the deep learning network?
In most of the autonomous driving work using deep reinforcement learning, it doesn’t take power consumption into account. They usually use few layers of convolutional layers in their deep reinforcement learning. In this project, taking the power consumption into consideration, several things to reduce number of trainable parameters, like applying 1×1 convolutional layers, replacing filter size of 7 with 3 filter size of 3, will be applied to shrink the model size. So the model is small enough to run on SRAM, which will effectively reduce the power consumption also.
QUADEER, Ahmed Abdul
January 19, 2022 3:44 pm
The relation between niche, problem, and solution is not clear.
Thanks for the comment and I’m sorry about the ambiguity. I will revise my poster to make it clearer. The niche is about the practicality (how much computational power required) and the safety (accuracy of the autonomous vehicle). If the trained model is able to achieve high accuracy but demands a very huge amount of computational power, it cannot be deployed in real life situation because usually the vehicles’ computational hardware is not that powerful. So the problem is to balance the amount of computational power required and the performance. I’m proposing a solution of the deep reinforcement learning with some power optimization strategies in my fyp. The deep reinforcement learning is a good learner that is able to achieve better performance than that of supervised learning, combining with some power optimization strategies, it will be able to achieve a balance the amount of computational power required and the performance.
Nice project. May I ask what are the major differences between this project and the existing autopilot systems in term of the architecture of the deep learning network?
In most of the autonomous driving work using deep reinforcement learning, it doesn’t take power consumption into account. They usually use few layers of convolutional layers in their deep reinforcement learning. In this project, taking the power consumption into consideration, several things to reduce number of trainable parameters, like applying 1×1 convolutional layers, replacing filter size of 7 with 3 filter size of 3, will be applied to shrink the model size. So the model is small enough to run on SRAM, which will effectively reduce the power consumption also.
The relation between niche, problem, and solution is not clear.
Thanks for the comment and I’m sorry about the ambiguity. I will revise my poster to make it clearer. The niche is about the practicality (how much computational power required) and the safety (accuracy of the autonomous vehicle). If the trained model is able to achieve high accuracy but demands a very huge amount of computational power, it cannot be deployed in real life situation because usually the vehicles’ computational hardware is not that powerful. So the problem is to balance the amount of computational power required and the performance. I’m proposing a solution of the deep reinforcement learning with some power optimization strategies in my fyp. The deep reinforcement learning is a good learner that is able to achieve better performance than that of supervised learning, combining with some power optimization strategies, it will be able to achieve a balance the amount of computational power required and the performance.
Thanks for the explanation.
Why there is no reference paper?
Thanks for your comment. I will add it later.