📂Machine Learning

MNIST Database

Overview¹

$$ \includegraphics[height=20em]{https://upload.wikimedia.org/wikipedia/commons/2/27/MnistExamples.png} $$

The MNIST database refers to a dataset of digit handwriting from American high school students and Census Bureau employees. It is commonly known as [MNIST].

• Official Website

Description

This dataset is frequently used as an example for beginners in machine learning/deep learning. NIST originally collected handwritten data in the following format for the evaluation of character recognition technology for automated sorting of handwritten postal codes. Yann LeCun took this handwritten data from high school students and Census Bureau employees, processed it, and created the MNIST. The image size is 28 x 28, and it consists of 60,000 training sets and 10,000 test sets.

$$ \includegraphics[height=30em]{https://www.nist.gov/sites/default/files/styles/960_x_960_limit/public/images/2019/04/27/sd19.jpg?itok=oETq77cZ} $$

How to Use

Julia

In Julia, the MNIST dataset can be used with the machine learning dataset package MLDatasets.jl. By default, it loads the training set in Float32 type. There are options to change this. The available methods are as follows:

• dataset[i]: Returns a tuple of the i-th features and target.
• dataset[:]: Returns a tuple of all features and target.
• length(dataset): Returns the number of data.
• convert2image(dataset, i): Converts the i-th data into a grayscale image. The ImageShow.jl package is required.

julia> using MLDatasets

julia> train = MNIST()
dataset MNIST:
  metadata  =>    Dict{String, Any} with 3 entries
  split     =>    :train
  features  =>    28×28×60000 Array{Float32, 3}
  targets   =>    60000-element Vector{Int64}

julia> test = MNIST(Float64, :test)
dataset MNIST:
  metadata  =>    Dict{String, Any} with 3 entries
  split     =>    :test
  features  =>    28×28×10000 Array{Float64, 3}
  targets   =>    10000-element Vector{Int64}

julia> length(train), length(test)
(60000, 10000)

julia> using Plots
julia> using ImageShow

julia> train.targets[1]
5

julia> heatmap(convert2image(train, 1))

Since the labels are given as integers, one-hot encoding needs to be done separately.

julia> train.targets[1:5]
5-element Vector{Int64}:
 5
 0
 4
 1
 9

julia> using Flux
julia> Flux.onehotbatch(train.targets[1:5], 0:9)
10×5 OneHotMatrix(::Vector{UInt32}) with eltype Bool:
 ⋅  1  ⋅  ⋅  ⋅
 ⋅  ⋅  ⋅  1  ⋅
 ⋅  ⋅  ⋅  ⋅  ⋅
 ⋅  ⋅  ⋅  ⋅  ⋅
 ⋅  ⋅  1  ⋅  ⋅
 1  ⋅  ⋅  ⋅  ⋅
 ⋅  ⋅  ⋅  ⋅  ⋅
 ⋅  ⋅  ⋅  ⋅  ⋅
 ⋅  ⋅  ⋅  ⋅  ⋅
 ⋅  ⋅  ⋅  ⋅  1

• How to do one-hot encoding in Julia Flux
• How to implement MLP and train MNIST in Julia Flux

Environment

• OS: Windows11
• Version: Julia v1.8.2, MLDatasets v0.7.6, Plots v1.36.1, ImageShow v0.3.6, Flux v0.13.7