This document discusses a project to implement neural networks on FPGAs for smart vehicles. It outlines the division of labor between three students for phases of the project. The first phase involves studying design implementations for parts of a convolutional neural network like batch normalization and multiplication. The second phase is implementing the convolution dataflow and pipelining, and determining the processing element structure. The third phase focuses on optimization. A GANTT chart outlines the timeline. Potential strengths, weaknesses, opportunities and threats of the project are also listed.
Simple, Complex, and Compound Sentences Exercises.pdf
Neurons On Wheels - Work Organisation
1. FPGA based systems for
automotive: Division of
labor
FPGA implementation of Neural
Networks for smart vehicles
Matteo Fettizio: matteo.fettizio@mail.polimi.it
Amin Machkouk: amin.machkouk@mail.polimi.it
Soufiane Machkouk: soufiane.machkouk@mail.polimi.it
Sala seminari DEIB, edificio
20
17/04/2019
2. First phase
Matteo Fettizio Amin Machkouk
Processing element
Study of Design Study of
Batch normalization Shift bit-based molteplication
implementation implementation
intra-level pipelining weight product(DSP)
4. Second and third phase
Implementation of the convolution dataflow and pipelining
P.E Structure End of the second phase
Optimization Third phase
5. 15/1 4/2 24/2 16/3 5/4 25/4 15/5 4/6 24/6 14/7
Computation of accuracy and loss parameters
Study of quantization techniques
Quantization of CNN
Implementation of the convolution dataflow and
pipelining
P.E. Structure
Optimization
Report
GANTT
Duration(Days)
6. •Low power consumption
•High speed end-to-end control
•High level of parallelism
•Lost in accuracy for high speed computation
•The project requires a lot of knowledge on CNN
•Acquire tools and knowledge
•Use this project for a future driverless car in formula student
•Autonomous vehicles are a target for a lot of researchers
THREATS (–)
STRENGTHS (+)
WEAKNESSES (–)
OPPORTUNITIES (+)