Design of Self-Driving Car Using Multitask Network
Abstract
Self-driving car
capabilities are severely impeded by the lack of two prominent
neuro-psychological constructs in extant artificial intelligence – perception
and anticipation. This project aims at tackling the problem of anticipation by
teaching neural networks to acquire domain knowledge. Domain knowledge
acquisition is the Holy Grail that machine learning endeavors to acquire.
However, current machine learning approaches to self-driving limit the
complexity of neural networks to single and simple tasks. This project
demonstrates an improvement in the capabilities of extant driving mechanism
using multitask learning - a learning paradigm which utilizes a single neural network
with several shared and unshared layers to predict several separate and
related, yet statistically independent tasks. Multitask neural networks
facilitate the acquisition of domain knowledge much like the human brain, in
that they multiplex related tasks in a single domain to the same neural
pathways. This multiplexing not only reduces the number of learnable
parameters, but the additional tasks make the learning models robust to
overfitting. The project also endeavours to implement the multitask learning
paradigm in hardware. A scaled-down electric car, powered by DC motors, was
mounted with an on-board raspberry-Pi camera module. During the data collection
phase, the pi module was IoT enabled facilitating the collection of training
data. During the autonomous telemetry phase, this module exchanged images and
corresponding driving commands using a two-way TCP connection with a remote
server.
Country : India
1 Dr. P. Sathiya
Professor, Department of Electronics and Communication Engineering, Malla Reddy College of Engineering for Women, Hyderabad -500100, Telangana, India
Argyriou, A., Evgeniou, T. Pontil, M.
Mach Learn (2008) 73: 243.
https://doi.org/10.1007/s10994-007-5040-8
Guillaume Obozinski , Ben Taskar.
Multi-task feature selection (2006).
Sebastian Ruder. An Overview of
Multi-Task Learning in Deep Neural Networks
Z. Yang, Y. Zhang, J. Yu, J. Cai, J.
Luo, "End-to-end multi-modal multi- task vehicle control for self-driving
cars with visual perception" in CoRR, 2018.
Liebel L, Korner K, Auxiliary Tasks
in Multi-task Learning, 2018.
