microsoft / tensorwatch

Debugging, monitoring and visualization for Python Machine Learning and Data Science
MIT License
3.42k stars 362 forks source link

Question & Discussion #69

Open imvetri opened 4 years ago

imvetri commented 4 years ago

Introduction

This is about starting a discussion to give neuron visualisation more importance than data visualisation.

Background

There are lot of tools and techniques available to visualise a trained network, trained model etc but there are no good visualisation technique for neuron itself.

Concept

A neuron is a reflection of space and time that is sensed by us, and most importantly we visually see it. In case of artificial neural network we cannot visually represent the neuron (If there is a technique let me know).

Need for visual representation

Human intelligence works because the neurons can learn sense other neurons (electrically - via impulse) and visually (atopsy) and artificial neural network lacks that.

By visually able to represent a neuron, it opens up lot other learning models that can stitch itself.

Example

Below example uses t-SNE technique to visualise the data set

image

Below tool helps to visualise the transformations on data

https://lutzroeder.github.io/netron/?url=https://raw.githubusercontent.com/nsfw-filter/nsfw-filter/master/dist/models/model.json

Expected outcome

Below is a simulation of an impulse in a neuron.

image

Question

How to build or figure out a technique to have a visual model where a neuron is represented in terms of space and the data represented in terms of light.