Evolutionary Artificial Neural Networks with Unity

Project Owner

  • ICT TechLab

Project Credits

  • Jim Cook (ICT, TechLab)
  • Lydia Gu (ICT, TechLab)

Start Date

29 Nov 2017

Status

Completed 12 Dec 2017 (investigating only)

Problem Statement

University of Sydney will open a new course: COMP5329 Deep Learning in semester one, 2018. Without a doubt, Deep Learning is going to dominate the trend in 2018, building on from the 2017 trend, machine learning. 

TechLab starts to investigate some open-sourced deep learning projects and this is one of them: applying Evolutionary Artificial Neural Networks with Unity. So far, the investigation includes training the Neural Network by switching the activation functions (sigmoid/tanh/softsign functions).

TechLab did some enhancements on this GitHub open-source project and tried to investigate:

  1. Neural network modelling; learn how the Evolutionary Algorithm works.

  2. Is it possible to develop and train a neural network project using framework and languages other than TensorFlow and Python?

  3. Is it possible to use Unity for the data visualisation of machine learning / deep learning training?

Final Brief

The investigation in TechLab had met its goal. The original project had also been enhanced (the activation functions) based on the knowledge from Stanford CS231n course (http://cs231n.github.io/) and deeplearning.ai courses (https://www.coursera.org/learn/neural-networks-deep-learning).

Challenges & Learnings

  • C1. Knowledge and skill sets of Deep Learning and Unity hindered further exploration.
  • L1. Fill the gap between knowledge and practice by learning deep learning, participating related trainings (e.g. Sydney Informatics Hub trainings)
  • L2. To offer machine learning Capstone projects that bring in group work and knowledge sharing from more master students who are knowledgeable in this area.

Languages / Framework

.NET framework, C#, Unity

Links to Resources

Github Enterprise Repo

References:

[1] The SoftSign function as proposed by Xavier Glorot and Yoshua Bengio (2010): “Understanding the difficulty of training deep feedforward neural networks” http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.207.2059&rep=rep1&type=pdf

[2] CS231n - Convolutional Neural Networks for Visual Recognition: Architecture http://cs231n.github.io/neural-networks-1



Tags: