Introduction to Programming and Scientific Applications (Spring 2021)

Project I - Geometric medians

Motivation: Assume you want to build a drone tower from where to deliver parcels to costumers, and you know the destinations (points) where to deliver parcels. What is the optimal placement of the tower, if each drone can at most carry one parcel and needs to return to the tower between each delivery?

The geometric median of a point set S = { p₁, ..., p_n } in the plane is a point q (not necessarily a point in S), that minimizes the sum of distances to the points, i.e. minimizes ∑_i=1..n |p_i - q|, where |p - q| = ((p_x - q_x)² + (p_y - q_y)²)^0.5.

1. Create a function geometric_median that given a list of points computes the geometric median of the point set.
   Hint. Use the minimize function from the module scipy.optimize.
2. Make representative visualizations that show both the input point sets and the computed geometric medians. Examples should include point sets with two and three points, and a random point set.
3. Use the matplotlib plotting function contour to plot a contour map to illustrate that it is the correct minimum that mas been found (the z-value for a point (x, y) in the contour map is the sum of the distances from (x, y) to all input points).

Next we want to find two points q₁ and q₂ such that the perpendicular bisector of q₁ and q₂ partitions the point set into the points closest to q₁ and those closest to q₂. We want to find two points q₁ and q₂ such that the sum of distances to the nearest point of q₁ and q₂ is minimized, i.e. ∑_i=1..n min(|p_i - q₁|, |p_i - q₂|) is minimized.

To solve this problem one essentially has to consider all possible bisection lines, and to find the geometric median on both sides of the bisector, e.g. using the algorithm from question a). Assuming no three points are on a line, it is sufficient to consider all n(n-1)/2 bisector lines that go through two input points, and for each bisector line to consider the two cases where the two points on the line both are considered to be to the left or to the right of the line.

4. Create a function two_geometric_medians that give a list of points p₁, ..., p_n computes two points q₁ and q₂ minimizing ∑_i=1..n min(|p_i - q₁|, |p_i - q₂|). It can be assumed that no three points are on a line.
5. Visualize your solution, showing imput points, q₁ and q₂, and the perpendicular bisector between q₁ and q₂.
6. [optional] Extend your solution to handle the case where three or more points can be on a line. E.g. what is your solution if you have four points on a line?

Project II - Portfolio optimization

We consider the problem of choosing a long term stock portfolio, given a set of stocks and their price over some period under risk aversion parameter γ > 0.

Assume there are m stocks to be considered. The portfolio will be represented by a column vector w ∈ ℝ^m, such that ∑_i=1..m w_i = 1. If w_i > 0, you use a fraction w_i of your total money to buy the i'th stock, while w_i < 0 represent shorting that stock. In both cases we assume the stock is bought/shorted for the entire period.

Let p_j,i represent the price of the i'th stock at time step j. If there are n + 1 time steps, then p ∈ ℝ^(n+1)×m is a matrix.

We let r ∈ ℝ^n×m be the matrix, where r_j,i represents the fractional reward of stock i at time step j, i.e. r_j,i = (p_j+1,i − p_j,i) / p_j,i for 1 ≤ j ≤ n. By r_j we denote the j'th row of r, viewed as a column vector (r_j,1, ..., r_j,m).

We make the (unrealistic) assumption that we can model r by a random variable, distributed as a multivariate Gaussian, with estimated means

μ ≃ 1/n · ∑_j=1..n r_j

and estimated covariance matrix

Σ ≃ 1 / n · ∑_j=1..n [(r_j − μ)(r_j − μ)^T]

Note that μ_i and Σ_i,i are the estimated mean and variance for stock i.

The distribution of returns using some w is then

R_w = N(μ_w, σ_w²)
μ_w = w^Tμ
σ_w² = w^TΣw

Now, we want to maximize for a balance between high return μ_w and low risk σ_w². This leads to the following optimization problem, where we want to find the value w* of w maximizing the following expresion:

maximize     w^Tμ − γw^TΣw

subject to     ∑_i=1..m w_i = 1

where γ controls the balance between risk and return. A high value of γ indicate we are willing to take low risk and vise versa.

In this project you should find w* for different values of γ and using real stock values of your choice. The project consists of the following three questions.

1. We need a module for collecting stock values. For this you can use the module pandas-datareader. Using this you should write a function get_prices([stock₁, ..., stock_k], step_size, period), that returns a tuple (stocks, p), where p[j, i] represents the opening price of stock i at time step j and stocks[i] is the name of the i'th stock (adjust the arguments to get_prices to the data available at your data source). Make a plot of p, where each stock is labeled with its name, e.g. MSFT or GOGL. You should use at least five stocks.
2. Calculate r, μ and Σ using the formulas above and the p calculated in question (a). Plot the probability density function (pdf) of the return of each stock.
   Hint. The method norm.pdf from the module scipy.stats might become convenient.
3. Solve the optimization problem defined above for different values of gamma, e.g. gammas = (np.arange(10) / 5) + 1, and plot the pdf of each solution to a single plot with appropriate legends. Finally create a scatter plot of how w* changes as γ changes: For each value of γ plot the fraction of each stock in the portfolio.

Project III - Simulations of NMR spectra of proteins

A fingerprint of a protein can be obtained by a nuclear magnetic resonance (NMR) experiment with a given spectrometer frequency (input_MHz). The fingerprint will be a spectrum of intensities over the frequency range. In this project you should compute a simulated approximation of the NMR fingerprint of a protein. In particular you should try to reproduce Figure 5(b) in [1]. The approximate fingerprint of a protein should be computed based on data available for the chemical shifts of the H-atoms in the protein and coupling data for H-atoms inside amino acids. We will approximate the spectrum with the sum of a set of Lorentz lines.

Protein molecules consist of one or more chains of amino acid residues. In the mandatory part of this project we will only consider proteins with one chain. Each amino acid in such a chain is identified by a a unique Seq_ID, an integer starting with one for the first amino acid in the sequence. The amino type is identified by Comp_ID (three letter abbreviation). Each atom in an amino acid has again a unique Atom_ID, where the first character is the atom type (i.e. in this project always H).

This project consists of the below tasks. Please remember to split your code into logical units, e.g. by introducing functions for various (recurring) subtasks.

1. Download the files NFGAIL.csv, couplings.csv and 68_ubiquitin.csv.
2. The simplest Lorentz line is the function L(x) = 1 / (1 + x²) . Make a function Lorentz that given x returns L(x), where x can be either an integer, a float or a Numpy array. In the case of a Numpy array the function should be computed pointwise for each value. Plot the function for x ∈ [−10, 10] using matplotlib.pyplot (Hint: use numpy.linspace).
   The function L has a maximum in (0, 1). We let (x₀, height) denote the coordinates of the maximum of a Lorentz line. We let the width of a Lorentz line be the width of the peak at height height / 2, sometimes denoted full width half height (FWHH). The function L has width 2.
   Note that the area below L is π = ∫_{(−∞,+∞)} L(x) dx.
3. Generalize your function Lorentz to take (optional keyword) arguments for x₀, height and width. The resulting function should compute L_{x0, height, width}(x) = height / (1 + (2 · (x − x₀) / width)²) . Plot three Lorentz lines for the parameters (x₀, height, width) being (-5, 5, 1), (2, 2, 6), and (5, 3, 0.5) for x ∈ [-10,10]. Plot also the sum of the three curves. Note that the area below a general Lorentz line is π · height · width / 2.
4. Our basic assumption is that each atom in a molecule contributes approximately one Lorentz line to the spectra. We will not use the same Lorentz parameters for all atoms. The width will e.g. depend on the atom_id and possibly also on the amino acid the atom is part of. Make a function peak_width(amino, atom_id) that returns the width for an atom in a specic amino acide and having a specific atom_id. If atom_id is H the width is 6. If (amino, atom_id) is one of the following four pairs (ASN, HD21), (ASN, HD22), (GLN, HE21), (GLN, HE22) the width is 25. For all other atom_id's the width is 4.
5. We will denote a Lorentz line a peak and identify it by the triple (x₀, height, width). For an atom (amino, atom_id) with assigned chemical shift value x₀ (in Hz) we create a peak with width determined by peak_width and maximum value at (x₀, height), where height is chosen such that the area below the Lorentz line is π. Create a function atom_peak(amino, atom_id, Hz) that returns the triple for the corresponding peak, where x₀ = Hz = the chemical shift in Hz. Plot the peaks for the three atoms and assigned chemical shifts (PHE, H, 3481 Hz), (ASN, HD21, 3053 Hz), and (ILE, HA, 1673 Hz). Furthermore plot the sum of the three peaks.
6. Create a function read_molecule that reads a list of atoms for a single protein from a CSV-file, where each row stores the description of an atom in the protein (most of the columns will not be used in this project). You can e.g. store each atom as a dictionary mapping column names to row values (Hint: use zip) or read the file as a pandas data frame.
7. Read in the molecule description of the NFGAIL protein from the file NFGAIL.csv. For each of the 44 H-atoms create a peak assuming input_MHz = 400.13 MHz. The relevant columns are Atom_ID, Comp_ID (= amino acid), and Val. The column Val is the chemical shift in ppm, parts per million, but this value should be converted to Hz for calculations. To get the shift in Hz, multiply the Val value by input_MHz. Plot the sum of the peaks. Make the unit of the x-axis ppm (= Hz / input_MHz) and make the x-axis be decreasing from left to right.
   Note. In the protein descriptions there is no H-atom listed for the first amino acid, since this H-atom will not be visible in the NMR spectra as it is in fast exchange with water.
8. To get a more refined spectrum we will take couplings between atoms into account. Between two H-atoms A and B in an amino acid (i.e. two peaks) there can be a coupling with magnitude J_AB, in the following just denoted J, that influences the resulting spectrum. (Many other factors influence the spectrum, but we will happily ignore these in our simulations). The coupling between A and B causes both peaks to be split into two new peaks A_inner, A_outer, B_inner and B_outer, where B_inner is closer to A than B, and B_outer is further away from A than B. In the following we only consider B_inner and B_outer (A is handled symmetrically). The height of B_outer is smaller than the height of B_inner, the sum of their heights equals the height of B, and their width equals the width of B. Assume A and B have their maximum at x₀ = ν_A and x₀ = ν_B, respectively. Let Q = sqrt((ν_A − ν_B)² + J²)
   ν_m = (ν_A + ν_B) / 2
   σ = 1 if ν_A < ν_B,   −1 otherwise
   α_inner = (1 + J / Q) / 2
   α_outer = (1 − J / Q) / 2 . The points B_inner and B_outer are given by ν_Binner = ν_m + σ · (Q − J) / 2 ,     height_Binner = height_B · α_inner ,     width_Binner = width_B ,
   ν_Bouter = ν_m + σ · (Q + J) / 2 ,     height_Bouter = height_B · α_outer ,     width_Bouter = width_B .
   Make a function apply_coupling(A, B, J) that takes two peaks A and B and computes the effect of A on B, when the coupling has magnitude J. Returns the list of the two peaks B_inner and B_outer that B is split into. If J = 0 or x₀(A) = x₀(B), then only [B] is returned. Plot the four peaks resulting from the mutual coupling of two peaks A = (25, 1, 1) and B = (75, 1, 1) with J = 10.
9. If an atom has couplings with several atoms the computations become slightly more involved. Assume B has couplings with k atoms A₁, A₂, ..., A_k with magnitudes J₁, J₂, ..., J_k, respectively. In general the peak for B will be split into 2^k peaks. The basic idea is to start with the peak B. Applying the coupling of B and A₁ with magnitude J₁ on B results in a list L₁ with at most two peaks. Applying the coupling of B and A₂ with magnitude J₂ on each B' ∈ L₁ results in the list L₂ with at most four peaks. Applying A₃ on the peaks in L₂ results in eight peaks L₃, etc.
   Applying the coupling between A_i and B with magnitude J on a peak B' ∈ L_{i − 1} is done as applying A on B, except that the final computation of the points B'_inner and B'_outer are given by ν_B'inner = ν_B' − ν_B + ν_m + σ · (Q − J) / 2 ,     height_B'inner = height_B' · α_inner ,     width_B'inner = width_B
   ν_B'outer = ν_B' − ν_B + ν_m + σ · (Q + J) / 2 ,     height_B'outer = height_B' · α_outer ,     width_B'outer = width_B
   Extend your function apply_coupling to also be able to take a peak B' as an additional (optional) fourth argument. Note that for B' = B the function should return the same value as without the B' argument.
   Make a function apply_couplings([(A₁, J₁), (A₂, J₂), ..., (A_k, J_k)], B) to apply all couplings on B. Plot the four peaks and their sum resulting from applying A₁ = (3, 1, 1) and A₂ = (5, 1, 1) on B₁ = (9, 1, 1) with magnitude J₁ = 1 and J₂ = 2, respectively. Note that permuting the list of (A_i, J_i)s should result in the same set of peaks for B.
10. Make a function to read the file couplings.csv that for each of the 20 amino acids contains the coupling magnitudes among the H-atoms inside the amino acid.
11. Make a function split_comps that splits a list of atoms for a protein into a list of sublists, one sublist for each amino acid on the chain, i.e. split the list into sublists based on Seq_ID.
12. Make a function comp_peaks(comp, couplings, input_MHz) that computes the peaks for all atoms in a amino acid (comp) by taking the table of couplings (from couplings.csv) into acount.
13. Make a function protein_peaks(atoms, couplings, input_MHz) that takes a list of atoms for a protein, a table of couplings, and a frequency input_MHz and returns the resulting peaks by applying all couplings. Apply the function to the data from NFGAIL.csv with input_MHz = 400.13 MHz. Plot the sum of the resulting 542 peaks. Make the unit of the x-axis ppm (= Hz / input_MHz) and make the x-axis be decreasing from left to right. Your result should be a reproduction of 5(b) in [1].

Protein	Uncoupled peaks	Coupled peaks
NFGAIL	44	542
Ubiquitin	556	6805
ThrB12-DKP-insulin	329	3338

The following tasks are optional.

In the previous tasks you were given CSV-files with the description of proteins only containing one chain. The following tasks ask you to download protein descriptions from the Biological Magnetic Resonance Data Bank (BMRB, bmrb.wisc.edu) containing an arbitrary number of chains. In the data bank each data set for a molecule has a unique Entry_ID value. Each chain is identified by a unique Entity_ID. Each amino acid in a chain has a unique Seq_ID (and amino type Comp_ID) and each atom in an amino acid has again a unique Atom_ID.

14. Download the file Atom_chem_shift.csv (700+ MB) from www.bmrb.wisc.edu > Bulk access > CSV relational data tables. This file contains the atom shift data for all data sets on the web site.
15. Write a function csv_extract that given a value Entry_ID, from Atom_chem_shift.csv extracts all rows with value Entry_ID in column Entry_ID and outputs the resulting rows to a new CSV-file. Test this with Entry_ID = 68 and Entry_ID = 6203. The result for Entry_ID = 68 should be the 556 atoms as provided in the file 68_ubiquitin.csv (but with additional columns). The result for Entry_ID = 6203 should be a file with a header row plus 329 data rows (atoms) for ThrB12-DKP-insulin. (Hint: Avoid reading the whole table into memory. If you e.g. read the complete table using Pandas, the data will take up memory ≈ 3 × file size. Instead scan through the file using the csv module and only have the current line in memory. Scanning through the 700+ MB should be possible within in the order of a minute.)
16. Update the function split_comps to be able to handle more than one chain of amino acids. In particular each comp should now be identified not only by Seq_ID but by the pair (Entity_ID, Seq_id).
17. Extract data for ThrB12-DKP-insulin (Entry_ID = 6203) from Atom_chem_shift.csv and plot the resulting simulated spectrum for input_MHz = 500 MHz. Note that the data for ThrB12-DKP-insulin consists of two chains of 131 and 198 atoms, respectively. You can find the Entry_ID value for your favorite protein by searching the BMRB data base at www.bmrb.wisc.edu.

References

[1] Fast simulations of multidimensional NMR spectra of proteins and peptides. Thomas Vosegaard. Magnetic Resonance in Chemistry. 56(6), 438-448, 2018. DOI: 10.1002/mrc.4663.

Project IV - MNIST Image Classification

In this project we are going to create a very simple neural network (linear classifier) to identify the handwritten digits from the MNIST database - often considered the "Hello World" problem in neural networks. In this problem we are given grayscale images of size 28 × 28 showing handwritten digits and are going to classify them into the 10 classes 0 - 9, depending on the digit depicted in the image. The MNIST database consists of 60.000 images to train your network on and 10.000 images to test the quality (accuracy) of the resulting network. For all images the correct label 0 - 9 is part of the database. There exist many of-the-shelf modules for this problem in Python, e.g. Keras, TensorFlow, scikit, and PyTorch, but in this project we are going to build a solution using pure Python from scratch.

A good introduction to the topic are the following four videos from the YouTube channel by 3BLUE1BROWN: Neural Network (19:13), Gradient Descend (21:00), Backpropagation (13:53), and Backpropagation Calculus (10:17). The few mathematical equations required in this project for performing simple backpropagation are stated explicitly below.

You are not allowed to use NumPy, Keras, etc. in the questions below (if not stated otherwise).

The first group of tasks concerns reading the raw data and visualizing them.

1. From yann.lecun.com/exdb/mnist/ download the following four files:
   • t10k-images.idx3-ubyte.gz (1.6 MB)
   • t10k-labels.idx1-ubyte.gz (4.4 KB)
   • train-images.idx3-ubyte.gz (9.6 MB)
   • train-labels.idx1-ubyte.gz (28.3 KB)
2. Make a function read_labels(filename) to read a file containing labels (integers 0-9) in the format described under "FILE FORMATS FOR THE MNIST DATABASE". The function should return a list of integers. Test your method on the files t10k-labels.idx1-ubyte.gz and train-labels.idx1-ubyte.gz (the first five values of the 10.000 values in t10k-labels.idx1-ubyte.gz are [7, 2, 1, 0, 4]). The function should check if the magic number of the file is 2049.
   Hint: Open the files for reading in binary mode by providing open with the argument 'rb'. You can either uncompress the files using a program like 7zip, or work directly with the compressed files using the gzip module in Python. In particular gzip.open will be relevant. To convert 4 bytes to an integer int.from_bytes might become useful.
3. Make a function read_images(filename) to read a file containing MNIST images in the format described under "FILE FORMATS FOR THE MNIST DATABASE". Test your method on the files t10k-images.idx3-ubyte.gz and train-images.idx3-ubyte.gz. The function should return a three dimensional list of integers, such that images[image][row][column] is a pixel value (an integer in the range 0..255), and 0 ≤ row, column < 28 and 0 ≤ image < 10000 for t10k-images.idx3-ubyte.gz. The function should check if the magic number of the file is 2051.
4. Make a function plot_images(images, labels) to show a set of images and their corresponding labels as titles using imshow from matplotlib.pyplot. Show the first few images from t10k-images.idx3-ubyte.gz with their labels from t10k-labels.idx1-ubyte.gz as titles. Remember to select an appropriate colormap for imshow.

A linear classifier consists of a pair (A, b), where A is a weight matrix of size 784 × 10 and b is a bias vector of length 10. An image containing 28 × 28 pixels is viewed as a vector x of length 784 (= 28 * 28), where the pixel values are scaled to be floats in the range [0, 1]. In the following we denote this representation an image vector. The prediction by the classifier is computed as

a = xA + b,

where a = (a₀, ..., a₉), i.e. the result of the vector-matrix product xA, that results in a vector of length 10, followed by a vector-vector addition with b. The predicted class is the index i, such that a_i is the largest entry in a.

In the follow we will apply the cost measure mean squared error to evalutate how close the output a = xA + b of a network (A, b) is for an input x to the correct answer y, where we assume that y is the categorical vector of length 10 for the correct label, i.e. y_i = 1 if i = label, and 0 otherwise:

cost(a, y) = sum_i ((a_i − y_i)²) / 10

We use the mean squared error because is has an easy computable derivative, although better cost functions exist for this learning problem, e.g. softmax.

Below you will be asked to compute weights (A, b) using back propagation. To get started, a set of weights (A, b) is available as mnist_linear.weights. The weights were generated using the short program mnist_keras_linear.py using the neural networks API Keras running on top of Google's TensorFlow. The network achieves around 92% accuracy on the MNIST test set (you should not expect to reach this level, since this network is trained using the softmax cost function).

5. Optional: You should be able to reproduce a similar weight file (but not exactly the same) by runing the script, after pip installing tensorflow. This is not part of the project.

The second group of tasks is to load and save existing linear classifier networks and to evaluate their performance, together with various helper functions. In the following we assume the vector b to be represented by a standard Python list of floats and the matrix A to be represented by a list-of-lists of floats.

6. Write functions linear_load(file_name) and linear_save(file_name, network) to load and save a linear classifier network = (A, b) using JSON. Test your functions on mnist_linear.weights .
7. Write function image_to_vector(image) that converts an image (list-of-lists) with integer pixel values in the range [0, 255] to an image vector (list) with pixel values being floats in the range [0, 1].
8. Write functions for basic linear algebra add(U, V), sub(U, V), scalar_multiplication(scalar, V) multiply(V, M), transpose(M) where V and U are vectors and M is a matrix. Include assertions to check if the dimensions of the arguments to add and multiply fit.
9. Write a function mean_square_error(U, V) to compute the mean squared error between two vectors.
   Examples: mean_square_error([1,2,3,4], [3,1,3,2]) shoule return 2.25.
10. Write function a function argmax(V) that returns an index into the list V with maximal value (corresponding to numpy.argmax). E.g. argmax([6, 2, 7, 10, 5]) should return 3.
11. Implement a function categorical(label, classes=10) that takes a label from [0, 9] and returns a vector of length classes, with all entries being zero, except entry label that equals one. For an image with this label, the categorical vector is the expected ideal output of a perfect network for the image.
    Example: categorical(3) should return [0,0,0,1,0,0,0,0,0,0].
12. Write a function predict(network, image) that returns xA + b, given a network (A, b) and an image vector.
13. Create a function evaluate(network, images, labels) that given a list of image vectors and corresponding labels, returns the tuple (predictions, cost, accuracy), where predictions is a list of the predicted labels for the images, cost is the average of mean square errors over all input-output pairs, and accuracy the fraction of inputs where the predicted labels are correct. Apply this to the loaded network and the 10.000 test images in t10k-images. The accuracy should be around 92%, whereas the cost should be 230 (the cost is very bad since the network was trained to optimze the cost measure softmax).
    Hint. Use your argmax function to convert network output into a label prediction.
14. Extend plot_images to take an optional argument prediction that is a list of predicted labels for the images, and visualizes if the prediction is correct or wrong. Test it on a set of images from t10k-images and their correct labels from t10k-labels.
15. Column i of matrix A contains the (positive or negative) weight of each input pixel for class i, i.e. the contribution of the pixels towards the image showing the digit i. Use imshow to visualize each column (each column is a vector of length 784 that should be reshaped to an image of size 28 × 28).

The third group of tasks is to train your own linear classifier network, i.e. to compute a matrix A and a vector b.

16. Create function create_batches(values, batch_size) that partitions a list of values into batches of size batch_size, except for the last batch, that can be smaller. The list should be permuted before being cut into batches.
    Example: create_batches(list(range(7)), 3) should return [[3, 0, 1], [2, 5, 4], [6]].
17. Create a function update(network, images, labels) that updates the network network = (A, b) given a batch of n image vectors and corresponding output labels (performs one step of a stochastical gradient descend in the 784 * 10 + 10 = 7850 dimensional space where all entries of A and b are considered to be variables).
    For each input in the batch, we consider the tuple (x, a, y), where x is the image vector, a = xA + b the current network's output on input x, and y the corresponding categorical vector for the label. The biases b and weights A are updated as follows:

    b_j −= σ · (1 / n) · ∑_(x,a,y) 2 · (a_j − y_j) / 10
    A_ij −= σ · (1 / n) · ∑_(x,a,y) x_i · 2 · (a_j − y_j) / 10

    For this problem an appropriate value for the step size σ of the gradient descend is σ = 0.1.
    In the above equations 2 · (a_j −y_j) / 10 is the derivative of the cost function (mean squared error) wrt. to the output a_j, whereas x_i · 2 · (a_j − y_j) / 10 is the derivative of the cost function w.r.t. to A_ij — both for a specific image (x, a, y).
18. Create a function learn(images, labels, epochs, batch_size) to train an initially random network on a set of image vectors and labels. First initialize the network to contain random weights: each value of b to be a uniform random value in [0, 1], and each value in A to be a uniform random value in [0, 1 / 784]. Then perform epochs epochs, each epoch consiting of partitioning the input into batches of batch_size images, and calling update with each of the batches. Try running your learning function with epochs=5 and batch_size=100 on the MNIST training set train-images and train-labels.
    Hint. The above computation can take a long time, so print regularly a status of the current progress of the network learning, e.g. by evaluating the network on (a subset of) the test images t10k-images. Regularly save the best network seen so far.

Here are some additional optional tasks. Feel free to come up with your own (other networks, other optimization strategies, other loss functions, ...).

19. Optional. Instead of using the mean squared error as the cost function try to use the categorical cross entropy (see e.g. this blog): On output a where the expected output is the categorical vector y, the categorical cross entropy is defined as CE(y, softmax(a)), where softmax(a)_i = e^a_i / (∑_j e^a_j) and the cross entropy is defined as CE(y, â) = − ∑_i (y_i · log â_i). In update the derivative of the cost function w.r.t. output a_j should be replaced by e^a_j /(∑_k e^a_k) − y_j.
    Note. softmax(a) is a vector with the same length as a with values having the same relative order as in a, but elements are scalled so that softmax(a)_i ∈ ]0,1[ and 1 = ∑_i softmax(a)_i. Furthermore, since y is categorical with y_i = 1 for exactly one i, CE(y, softmax(a)) = log(∑_j e^a_j) − a_i.
20. Optional. Visualize the changing weights, cost, and accuracy during the learning.
    Hint. You can use matplotlib.animation.FuncAnimation, and let the provided function apply one batch of training data to the network for each call.
21. Optional: Redo the above exercises in Numpy. Create a generic method for reading IDX files into NumPy arrays based on the specification "THE IDX FILE FORMAT". Data can be read from a file directly into a NumPy array using numpy.fromfile and an appropriate dtype.
    Hint. np.argmax(test_images.reshape(10000, 28 * 28) @ A + b, axis=1) computes the predictions for all tests images, if they are all in one NumPy array with shape (10000, 28, 28).
22. Optional: Compare your pure Python solution with your Numpy implementation (if you did the above optional task) and/or the solution using Keras, e.g. on running time, accuracy achieved, epochs.
23. Optional: Try to take a picture of your own handwritten letters and see if your program can classify your digits. It is important that you preprocess your images to the same nomalized format as the original MNIST data: Images should be 28 × 28 pixels where each pixel is represented by an 8-bit greyscale value where 255 is black and 0 is white. The center of mass should be in the center of the image. In the test data all images were first scaled to fit in a 20 × 20 box, and then padded with eight rows and columns with zeros to make the center of mass the center of the image, see yann.lecun.com/exdb/mnist.
    Hint: PIL.Image.resize from the Pillow (Python Imaging Library) might be usefull. Remember to set the resampling filter to BILINEAR.

Project V - Open topic

You can also define your own project. There are some aspects that needs to be addressed.

• Some minimal amount of Python 3 code must be produced.
• Use at least two different Python modules.
• Some data must be visualized, e.g using matplotlib, but any visualization module or domain specific library, like for chemistry, can be used.

You need to get your project description approved by sending an e-mail to gerth@cs.au.dk.