Burn MNIST exploration

Use to train a simplistic 768-16-16-10 DNN mentioned in 3blue1brown Neural networks series on MNIST dataset and see what happens.

(Just following the Burn book really)

Results

Reached 91.9% accuracy :)

======================== Learner Summary ========================

Model:

Model {
  activation: Relu
  hidden1: Linear {d_input: 768, d_output: 16, bias: true, params: 12304}
  hidden2: Linear {d_input: 16, d_output: 10, bias: true, params: 170}
  output: Linear {d_input: 10, d_output: 10, bias: true, params: 110}
  params: 12584
}

Total Epochs: 10

Split	Metric	Min.	Epoch	Max.	Epoch
Train	Accuracy	53.035	1	91.817	10
Train	Loss	0.285	10	1.536	1
Valid	Accuracy	78.310	1	91.940	10
Valid	Loss	0.282	10	0.897	1