How to print the info:

sinfo -o "%9n %9P %.4c %.6m %.8d %14f %G"

How to format with xsv:

sinfo -o "%n|%P|%c|%m|%d|%f|%G" | xsv table --delimiter "|"

I consider it rather fortunate since it allows me to clean up, as demonstrated below.

Placement of the goods is a confounder.

hit-and-run

Configuration that works in Thunderbird in English in Ubuntu:

• Incoming:
  • Authentication: Normal password
  • Username: bartefil
• Outgoing:
  • Authentication: Normal password
  • Username: bartefil

Slovenské křesťanské rádio. Viz lumen.sk.

y = (x - loc) / scale
return c * np.power(y, -c-1) * np.power(1 + np.power(y, -c), -2) / scale

Emulation General Wiki contains comparisons of emulators of various platforms.

Student Wiki

exploration

\(p_\theta(x)\)

Estimated probability that the label of x is y

Musí být více než 75 % z osobních nákladů vyplaceno na stipendia studentů, nebo 75 % z osobních nákladů akademiků, nebo 75 % z celkových nákladů na projekt?

Recommended literature: Deep Learning Book

The higher mu is, the more closely aligned the true gradient and g must be.

[close on average]

learning rate

Gradient estimate

\(\lambda_i > 0, \sum \lambda_i = 1: \log \sum \lambda_i x_i > \sum \lambda_i \log x_i\)

Starting with m, all the following approximations are strictly contained within epsilon-wide sleeve around the true L.

\(z^j = (x^j, y^j)\)

The more training data we have, the closer we get to the true optimum.

\(p(y|x) = \frac{p(x, y)}{p(x)} = \frac{p(x|y) p(y)}{p(x)} = ... \)

We assume a certain model architecture. We need to estimate the parameters of the model.

Unknown parameters:

• Prior: alpha
• Posterior: mu_0, mu_1, V

Example of discriminative learning version 2

Example of discriminative learning version 1

Example of generative learning

Example: Softmax layer for classification - estimates class probabilities

Correction: min

posterior

Why "likelihood" rather than "probability"? If the domain is real valued, p is a PDF and summing over p values does not correspond to probability.

What does the network output?

1. Value estimate (required for this assignment)
2. Policy (to be introduced in later lecture)

This expression should be negated so that the loss increases as \(p_k\) decreases.

Assumption: y is a linear function of x with Gaussian noise with the same standard deviation (sigma) in each dimension.

Maximum likelihood estimation

Lecture slides: https://cw.fel.cvut.cz/wiki/_media/courses/be4m33ssu/anns_ws19.pdf

Recommended reading:

• https://en.wikipedia.org/wiki/Matrix_multiplication
• https://en.wikipedia.org/wiki/Invertible_matrix
• https://en.wikipedia.org/wiki/Matrix_calculus
• https://en.wikipedia.org/wiki/Derivative
• https://en.wikipedia.org/wiki/Partial_derivative

JD: The corresponding backward message may be one of the exam assignments.

Wikipedia

the universal approximation theorem states that a feed-forward network with a single hidden layer containing a finite number of neurons can approximate continuous functions on compact subsets of Rn, under mild assumptions on the activation function.

Skutečné údaje:

• Minimální stipendium za měsíc: 9 000 Kč. Zdroj: Příkaz rektora č. 11/2019
• Minimální stipendium za kalendářní rok: 12 x 12 000 Kč. Zdroj: Stipendijní řád - článek 6:1-3

Python: Use the library svmlight. Consider using sklrean.svm.

Solution consists of:

• Source code (Python or Matlab)
• PDF with answers to questions (and nothing else)

Alternative from Xavier: \(\frac{2}{n_{in} + n_{out}}\)

Is it safe to assume that x_i are independent across i?

Independent across i

We may need to normalize the input.

Learning rate may be further specialized e.g. for each layer.

We move in the opposite direction of the gradient.

Class confidence distribution

Proof: Seminar 3, Assignment 4

1. \(R(h_m) - R_{T^m}(h_m) \leq \sup_{h \in H} |R(h) - R_{T^m}(h)|\) - trivial
2. \(R_{T^m}(h_H) - R(h_H) \leq \sup_{h \in H} |R(h) - R_{T^m}(h)|\) - easy to see

\(R_{T^m}(h_m) \leq R_{T^m}(h_H)\) because \(h_m \in \argmin_{h \in H} R_{T^m}(h)\).

Neural networks need not converge to the global minimum. \(R(h_m)\) need not equal \(R(h_H)\).

\(D_{VC}(H) = max_n . \exists x \in X^n . \forall y \subseteq x . \exists h \in H . h(x) = y\)

We assume 0-1 misclassification loss.

Space of feature vectors

and bounded loss image (bounded by [a, b])

"Uniform" because we care about all strategies in H.

Equivalently:

\(\forall \delta > 0, \epsilon > 0. \exists m_0 \in N. \forall m \geq m_0. P(R(h_m) - R(h_H) \geq \epsilon) < \delta\)

In general:

\(lim_{m \rightarrow \infty} f(m) = 0 \Leftrightarrow \forall \delta > 0. \exists m_0 \in N. \forall m \geq m_0. f(m) < \delta\)

How to reduce estimation error?

• Increase number of samples m.
• Improve the learning algorithm to utilize T^m more efficiently.
• Shrink H to exclude strategies better than h_m.

Enlarging H typically leads to worse generalization, so estimation error increases.

R(h_m) is a random variable, because h_m is a random variable, because T^m is a random variable.

All functions \(h: X \rightarrow Y\)

How can we reduce approximation error?

Enlarge H so that it includes better rules.

An empirical risk minimization (ERM) algorithm returns \(h_m \in H\) that minimizes \(R_{T^m}(h)\).

• Excess error: \(R(h_m) - R^*\)
  • Estimation error: \(R(h_m) - R(h_H)\)
  • Approximation error: \(R(h_H) - R^*\)

An algorithm is statistically consistent in \(H\) iff \(\forall \epsilon > 0: \lim_{m \rightarrow \infty} P(R(h_m) - R(h_H) \geq \epsilon) = 0\).

The hypothesis space \(H\) satisfies the uniform law of large numbers (ULLN) iff \(\forall \epsilon > 0: \lim_{m \rightarrow \infty} P(\sup_{h \in H} |R(h) - R_{T^m}(h)| \geq \epsilon) = 0\).

Theorem 1: If \(H\) satisfies ULLN then ERM is statistically consistent in \(H\).

Corollary 1: The ULLN is satisfied for a finite \(H\).

Generalization bound for finite hypothesis space \(H\) with \(Im(l) \subseteq [a, b]\): \(P(\max_{h \in H} |R_{T^m}(h) - R(h)| < \epsilon) \geq 1 - 2 |H| \exp(-\frac{2 m \epsilon^2}{(b - a)^2})\)

Boris:

We usually publish the slides 1 hour before the lecture.

The environment adheres to reward shaping. Each state has a hidden value and reward of an executed action is always the difference between the values of the previous and the current state. (?)

We can submit one or more Python source files (modules) with the .py file extension.

Most likely this is too short period to perform training in ReCodEx.

Episode terminates after 1000 steps.

There are 500 steps in each episode.

Is gamma equal to 1.0 in evaluation?

Assignments: Best solution counts. There is no penalty for incorrect solution submission. The students can ask for increasing the submission count limit (50 by default).

Train either in ReCodEx or offline.

If I can't get it running in an hour, I can write to MS and ask for advice.

Generalized states: R^4:

1. Cart position
2. Cart velocity
3. Pole angle
4. Pole angular velocity

For this task the states are discretized by binning (8 bins in each dimension; 8^4 states in total).

We improve the policy steps times.

After iterations evaluations, we improve the policy once.

Average value per assignment: 11 points

I can ask for a justified deadline extension before the deadline passes.

Given a valuation and a state, returns the expected return provided we follow the action that maximizes the next reward plus the valuated value of the next state.

Return at time \(t\)

The smaller the discount factor, the smaller is the relative weight of the far future rewards.

It may make sense to decrease epsilon gradually while training to explore less and exploit more (refine the policy for the successful runs).

Can we make the policy stochastic and choose each action with a probability that scales with its expected return?

Under what conditions do we get stuck in a local optimum?

Český ústav robotiky a kybernetiky

Finite C implies forces w to be in a hypersphere of finite radius R that depends on C.

Algorithm used: Matlab: Interior point

https://www.mathworks.com/help/optim/ug/fmincon-interior-point-algorithm-with-analytic-hessian.html (?)

Exponential moving average

\((1 - \alpha) Q_n + \alpha R_n\)

New estimate for this machine is old estimate plus relative current reward divided by number of times we tried this machine.

Cf. gradient descent

\epsilon: Rate of exploration

Non-differentiable

Unfair interface advantage

Very high computational demands

Can beat the previously best program after 3h of playing

Nickname: master

Online game server, the program did not reveal it was a bot

Why is go interesting?

• High branching factor
• Difficult to write a good state evaluation heuristic

Plays all the games

Questionable information value because some games can be exploited a lot

A separate agent trained for each game

1959

Learning in animals

One input may have more than one label with non-zero probabilities.

conv3-64: receptive field 3x3, stride 1, 64 output channels

To resolve boundary: pad with zeros

MP: max pooling layer. Window 2x2, stride 2, aggregate with max.

Deep layers: more channels, smaller resolution, large transitive receptive field, more abstract features

Surrogate loss function (or objective function) for 0-1 loss

Why? So that it is differentiable.

If we used 0-1 loss, the empirical risk would be the error rate.

To choose the class, we take the class with the highest predicted probability.

Optimization problem

Filip: Is the estimation based on Hoeffding inequality commonly used in practice (research)?

A: Yes.

Strength: Valid irrespective of the underlying distribution

What distribution maximizes the error?

This equality binds the variables \epsilon, \gamma, l and (b-a).

The variables are bounded interval by an interval \([a, b]\).

• Left graph
  • Red line: True expectation (3.5)
  • Blue band: \epsilon band around assumed expectation
• Right graph:
  • Red curve: Portion of experiments that estimate a mean outside the blue band in the left graph
  • Blue curve: Upper bound of the empirical established by Hoeffding inequality

3 interdependent variables:

• \(\epsilon\)
• \(l\)
• \(\delta\)

3 possible tasks:

1. Given l and \delta, minimize \epsilon.
2. Given \epsilon and l, maximize \delta.
3. Given \epsilon and \delta, minimize l.

Confidence interval

BF:

NP complete or NP hard

Theoretical assignments every other week published one week in advance.

No evaluation, just discussion.

90 minutes. Assignments similar to those discussed in the labs.

The course will run irrespective of the number of students. The lecture format will only be held if at least approx. 5 students come consistently.

Boris:

You should know this already.

Note that the validation set must still be labeled.

Note that this only works for classification.

Is this infimum?

Example: For 0-1 loss, using Chebyshev inequality, using the loss bound of 1, we get the bound: $$\frac{1}{\epsilon^2 m}$$

The evaluation distribution must match the application distribution. In other words, we evaluate the inference with respect to p.

independent identically distributed

In case of 0-1 loss in classification, this is the error rate of h.

Expectation

The distribution of (x, y) will be specified later.

We assume that x and y are interdependent.

Simplest loss for classification: 0-1 loss: l(a,b) = 1 iff a != b

possibly a vector

Capacity: 30 people

Boris:

Don't hesitate to contact us.

Don't hesitate to ask questions during the lectures.

Immediately after each lecture we will go to the cafeteria where you can contact us privately.

Each task: Around 10 points plus up to 3 bonus points

Boris:

a week ahead

Introductory lab session. Neither theoretical, nor practical.

To have a personal certificate generated by CIIRC, follow the instructions in the wiki.

Credentials: CTU username and CIIRC password

Which version of Ubuntu is this guide intended for?

Does this mean that modification of Vampire for research purposes is prohibited?

According to the paper (Cohen 1998), maximizing the permutation score is an NP-hard problem. The paper proposes a 2-approximation greedy algorithm.

https://www.quora.com/What-is-the-meaning-of-latent-space

Bratr Marka Rosy z GoodAI

The inverse is true. The text contains "U -> U^k".

#ProblemGenerators

Moc emocí, málo informací a argumentů.

Myslím, že tento postoj je zároveň součástí ideologie autora tohoto textu.

Apel na emoce

Tohle mi připomíná manipulativní heslo Parlamentních listů:

Nikdo nám nediktuje, o čem smíme psát.

Kritická analýza: Jeden svět na školách

Zatím se mi zdá, že qTube vlastní níže podepsaný autor tohoto textu.

Chybí zdroje.

Co když vás odsoudím za manipulativní rétoriku?

Color LaserJet cm6040 MFP

Viz screenshot v bodu 6 níže.

Ubuntu 18: Settings -> Devices -> Printers -> Additional Printer Settings...

Zdroj: Ask Ubuntu

Ubuntu 18: Add

This reminds me of the numerous shots of Al Gore at an airport or on an airplane in the 2006 environmentalist film An Inconvenient Truth.

Katedra počítačů

Katedra kybernetiky

Satisfiability modulo theories

Such property is implicit: = @ andrew

A heated mixture of coffee and any syrup is still coffee.

A heated mixture is hot!

A heated mix is the application of heat to the mixture of a beverage with a syrup.

A heated mix is the application of heat to the mixture of a beverage with a syrup.

A mixture of a beverage with a syrup is a beverage.

A heated beverage is a beverage.

Coffee is a beverage.

IMDB list: Vatican Best Films List

Judging from the context, the author meant "NP hard".

one item to one character

https://mddcarpet.github.io/

https://www.topcoder.com/challenges/30081256

shangjingbo:

The density in the current test cases is no more than 0.2.

According to songzy12, it is necessary to call the function at least once to get a score. See discussion in the forum.

There may be multiple edges between any pair of nodes.

shangjingbo:

I think you should accept multiple edges between the same pair of nodes.

sudo0124:

"Total execution time(ms)" in the email is the time we are using for scoring that counts towards 180 sec. limit.

-- Meaning "Total execution time(ms)" in the email is the T, and "DA API wait time" is already excluded from there.

https://www.zakonyprolidi.cz/cs/2017-341

Source code: github.com/filipbartek/hashcode2019qualification

Moreover, this constraint ensures that each column has at least one number of each in \(K\). The constraint ensures that each column contains each number in \(K\) exactly once.

Instrukce v této sekci jsou platné pouze pro Windows. Pokud máte Linux, následujte instrukce v jednom z následujících návodů:

• How to Create a Bootable Windows 10 USB in Linux
• Create a Bootable Windows 10 USB in Linux With Ubuntu/Debian GUI

Why do you think so?

How exactly would that help in this case?

maximální výše PvN činí 0,58násobek průměrné mzdy v národním hospodářství za první až třetí čtvrtletí kalendářního roku předcházející kalendářnímu roku, ve kterém byla podána žádost o PvN.

Zdroj: Základní poučení

Průměrná mzda v národním hospodářství v roce 2017: 28 761 Kč

Zdroj: Sdělení MPSV

Maximum v roce 2018: 0,58 * 28 761 = 16 681,38 Kč

200,00 Kč

Zdroj: https://www.ceskaposta.cz/documents/10180/282453/celni-deklarace.pdf/d8ca16ed-4969-4d68-aa4c-d4ec261b4e91

%APPDATA%\Dr. Memory

Why should I add myself to "audio" group?

WikiWikiWeb

The principle can also be applied to source code files.

singular

It seems that the edge probe is a special case of the value probe.

More specifically, you'll need to know up front all the places you want your patch commit applied into so that you can determine where to start the patch branch from.

This is an important assumption.

The way I understand it, cherry picking is intended to be used in case of unanticipated migration of code.

The patch branch should start at a common ancestor of all the target branches. Since the broken code is in all the target branches, it must be in at least one of their common ancestors.

If we don't start from a common ancestor and merge the patch into all the target branches, at least one of the target branches will get some extra change along with the patch.

Is it really possible to get an "undefined variable" compilation error from a function defined in a DLL?

First of all, the calls to the DLL are only resolved by the linker. Second, I don't think the linker analyzes the definitions of the functions at all. Third, I don't think there are any "variables" in native binary code; there are just memory addresses and registers in play.

Komu?

See Wikipedia: Convention over configuration.

Also try the command where in Windows command prompt. It does not only run in PowerShell.

Do not give up personal interaction in favor of a tool-supported process. "Když tomu nerozumím, tak se jdu zeptat."

On my computer the device with the unused space is /dev/sdb. You can see it in the graphical utility Disks.

Why?

VMware Workstation Player 12:

/dev/sr0 on /media/ubuntu/VMware Tools type iso9660 (ro,nosuid,nodev,relatime,uid=1000,gid=1000,iocharset=utf8,mode=0400,dmode=0500,uhelper=udisks2)

What does this mean?