4 - [Learn numpy](https://numpy.org/learn/)
5 - [MNIST database of handwritten digits](http://yann.lecun.com/exdb/mnist/)
6 - [Neuron](https://en.wikipedia.org/wiki/Artificial_neuron)
7 - [Perceptron](https://en.wikipedia.org/wiki/Perceptron)
8 - [Backpropagation](https://en.wikipedia.org/wiki/Backpropagation)
9 - [Understanding & Creating Neural Networks with Computational Graphs from Scratch](https://www.kdnuggets.com/2019/08/numpy-neural-networks-computational-graphs.html)
10 - [3Blue1Brown video series](https://www.youtube.com/playlist?list=PLZHQObOWTQDNU6R1_67000Dx_ZCJB-3pi)
12 Too high-level for first-time learning, but apparently very abstract and powerful for real-life:
13 - [keras](https://keras.io/)
14 - [tutorial how to recognize handwriting with keras/tensorflow](https://data-flair.training/blogs/python-deep-learning-project-handwritten-digit-recognition/)
18 sudo dnf install -y python3-numpy python3-matplotlib
22 - Do the [NumPy quickstart tutorial](https://numpy.org/devdocs/user/quickstart.html); example:
26 import matplotlib.pyplot as plt
27 grad = np.linspace(0,1,10000).reshape(100,100)
28 plt.imshow(grad, cmap='gray')
31 plt.imshow(np.sin(np.linspace(0,10000,10000)).reshape(100,100) ** 2, cmap='gray')
32 # non-blocking does not work with QT_QPA_PLATFORM=wayland
37 - Get the handwritten digits training data with `./download-mnist.sh`
39 - Read the MNIST database into numpy arrays with `./read_display_mnist.py`. Plot the first ten images and show their labels, to make sure the data makes sense:
41 
43 - Define the structure of the neural network: two hidden layers with parametrizable sizes. Initialize weights and biases randomly. This gives totally random classifications of course, but at least makes sure that the data structures and computations work:
47 output vector of first image: [ 0. 52766.88424917 0. 0.
48 14840.28619491 14164.62850135 0. 7011.882333
50 classification of first image: 1 with confidence 52766.88424917019; real label 5
51 correctly recognized images after initialization: 10.076666666666668%
54 - Add backpropagation algorithm and run a first training round. This is slow, as expected:
57 output vector of first image: [ 0. 52766.88424917 0. 0.
58 14840.28619491 14164.62850135 0. 7011.882333
60 classification of first image: 1 with confidence 52766.88424917019; real label 5
61 correctly recognized images after initialization: 10.076666666666668%
62 round #0 of learning...
63 ./train.py:18: RuntimeWarning: overflow encountered in exp
64 return 1 / (1 + np.exp(-x))
65 correctly recognized images: 14.211666666666666%
72 - This is way too slow. I found an [interesting approach](https://www.kdnuggets.com/2019/08/numpy-neural-networks-computational-graphs.html) that harnesses the power of numpy by doing the computations for lots of images in parallel, instead of spending a lot of time in Python on iterating over tens of thousands of examples. Now the accuracy computation takes only negligible time instead of 6 seconds, and each round of training takes less than a second:
75 output vector of first image: [0.51452796 0.49736819 0.51415083 0.50027547 0.48447025 0.49759904
76 0.52621162 0.48671402 0.517606 0.50214569]
77 classification of first image: 6 with confidence 0.526211616929459; real label 7
78 correctly recognized images after initialization: 7.75%
79 cost after training round 0: 1.0462266880961681
81 cost after training round 99: 0.4499245817840479
82 correctly recognized images after training: 11.35%
89 - Poor recognition quality after 100 iterations, as the network structure is apparently inappropriate. Having only 16 neurons in the first hidden layer makes the network not able to "see enough details" in the input. So let's use 128 neurons in the first hidden layer, and drop the second layer (it only seems to make things worse for me). Naturally a single training round takes much longer now, but voilĂ , after only 20 learning iterations it's already quite respectable:
92 correctly recognized images after initialization: 9.8%
93 cost after training round 0: 0.44518003660592853
95 cost after training round 19: 0.10783488337150668
96 correctly recognized images after training: 89.09%
102 - And after 100 iterations, the accuracy improves even more, and the classification of the first test image looks reasonable:
104 cost after training round 99: 0.043068345296584126
105 correctly recognized images after training: 94.17%
107 output vector of first image: [1.11064478e-02 5.59058012e-03 5.40483856e-02 7.93664914e-02
108 2.22662031e-03 3.50355065e-03 2.57506703e-04 9.60761429e-01
109 2.68869803e-03 5.26559410e-03]