Neural Network Playground

Tensorflow playground

In the context of a call for projects from a large European energy group, I developed a platform to raise awareness of the energy challenges of Deep Learning model training among AI players.

Principle and objective

The user plays the role of an engineer who has to port an algorithm from R&D and based on a neural network. He has the database and the global structure of the neural network, as well as objectives in the form of value ranges.

How to play ?

To achieve its objectives, it can play on 4 variables:

  • The choice of hardware platform defined by the type of processor, amount of memory, etc. The platform has a cost and of course will constrain performance.
  • A scale parameter that defines the "depth" of each layer.
  • A parameter on the resolution of the input images (from the camera).
  • The accuracy of the representation of the numbers in the network to make the calculations.

To win, it must find a configuration of these 4 parameters that fall within the ranges of the following 4 performance criteria:

  • FPS: number of frames per second processed by the onboard system.
  • Energy: Frugality is one of the three pillars of trusted AI.
  • Cost: A product only makes sense if it is economically viable.
  • Clarification: the overall performance of the network in the sense of the effectiveness in fulfilling its task.


The game was developed in Vue.JS (including HTML/CSS and JS). The computation times and costs according to the platforms used were obtained by making API calls to the group's hardware platform and using different neural network architectures, so the game faithfully represents reality.