Impressions and Links from
ML Prague 2019

I had the great pleasure of taking part in ML Prague 2019 (The practical conference about ML, AI and Deep Learning applications. February 22 - 24, 2019. Prague).

Below you will find impressions from the conference, and links for further reading.

The ML Prague 2019 conference was held in the beautiful city of Prague (Czech Republic).

Tried to follow as many talks as possible. But, well, these notes are, of course, in no way, shape or form complete...
Rather, these notes were written on conference nights, as my way of keeping track of the events that I attended at the conference. And as a way of storing links and references for future reference.

But enough disclaimers, below, you'll find impressions and links from some of the conference talks and seminars, including links for further reading.

Great stuff indeed. And much (ML & AI stuff) to look forward to in the coming years!

1. Introduction.

The venue for ML Prague 2019 was the impressive Rudolfinum, in Prague,
associated with music and art since its opening in 1885.

1.1. Page Overview.
- Presentations, Keynotes and Workshops.

Below, in section (2 - 4), you will find impressions and links from presentations that I followed Saturday and Sunday. As well as the workshop that I followed Friday. In section (5) I have included post-conference impressions from Prague. Additional Prague pictures follow in section (6).

Please notice: These notes don't do justice to the often brilliant presentations that initiated them!
So, please read the original presentations to avoid any distortions ...

2. Impressions from the Workshop, Friday.
February 22nd.

2.1. Workshop: Deep Learning for Text Processing.

The workshop venue was the captivating CEVRO Institut, Jungmannova 28/17, in Prague.

The teachers for ''Deep Learning for Text Processing'' workshop at Machine Learning Prague 2018 was Petr Baudis and Martin Holecek from Rossum. Where Rossum specializes in building Artificial Intelligence to understand documents (using neural network architectures for information extraction).

Our first experiment would be an ''understanding'' task. Where we would look at sentiment classification of movie reviews (The review sentiments, we worked with, were classified according to a good to bad axis).

Clearly, teaching AI (and teaching the IT community to integrate AI in existing solutions) isn't easy.

Still, the sessions quickly turned interesting, in Petr & Martin's capable hands, as we digged into to their pre-prepared ''Sentiment Analysis on IMDB Reviews'' code examples (See their Github page).

First, Bag of Word- and Word Embedding- models were explained.

Including Word2Vec (I.e. - a group of models that are used to produce word embeddings).
Where a given corpus of text is used to create a vector space, where:

Word vectors are positioned in the vector space such that words that share common contexts in the corpus are located in close proximity to one another in the space.

Various convolutional neural network models were then introduced in order to classify texts.

Which moved us further into Natural Language Processing.

With text classification according to topics and sentiments. I.e. where we hope to identity topics in text segments, along with the dominating sentiments in the segments.

Combining several techniques turned out to be a smart move, which gave pretty good results, even without coding that much (See the GitHub page).

Adversarial methods were also shortly introduced (See Adversarial training methods for semi-supervised classification). With the indication that such models might lead us to even better results.

All in all, a super inspiring workshop!
(Followed by equally interesting ''Analyzing Social Media Data''- and ''Artifical Intelligence Ethics''- workshops).

For more about Natural Language Processing, see e.g. MlCon 2018 (Section 3.1. Text classification and NLP on Star Wars characters), and below (Sections 3.6 & 3.7) ...

3. Impressions from Saturday.
February 23rd.

3.1. Why inverse reinforcement learning is impossible, and why we can do it anyway.

Stuart Armstrong is a ''Alexander Tamas Fellow in Artificial Intelligence and Machine Learning'' at Oxford University with an interest in the safety and possibilities of Artificial Intelligence.

Basically, inverse reinforcement learning (IRL) is about learning from humans.
I.e. learning an agents objectives, values, or rewards by observing the agents behavior.

The goal of the learning agent is to find a reward function from the expert demonstrations that could explain the expert behavior.

(Not mentioned in the talk..., but clearly... A capability machines need, as AI reaches super-human levels..., and need to understand what humans want, and (hopefully) tries to work towards these goals.
Certainly, I couldn't help thinking about things like Asimovs ''Law Zero''.
Simple: Take a set of human-generated data for a task, and extract an approximation for the reward function, for the task, ... Indeed, stuff like this used to be Science Fiction...).

The problem is, of course, that there might exist many different reward functions that the agent (we are looking at) might be attempting to maximize.
And for most real-world tasks, it is also a problem that humans have limited information about their world, and that their information about the world changes over time. Again, making it pretty hard to figure out what their reward function actually is...

Not to mention that:

Agents will often take long series of actions that generate negative utility for them in the moment in order to accomplish a long-term goal.

(See: Thinking Wires by Johannes Heidecke).

Stuart Armstrong also pointed out that people in different cultures tend to interpret certain behaviours differently. E.g. some people see certain behaviours as rational, while others might think that the same behaviours are irrational. Indeed, there are no end to the problems here, when we try to do IRL generally.

So, if we really want to know what is going on inside a humans head...
To understand e.g. why a human is playing a hand in poker the way he or she is (Wanting, either to win, or to lose. To win money, or to lose... E.g. Alice might fancy Bob, and think losing will boost Bobs ego, and make him fancy her...), we could use MRI (and directly access their brains...).
But, we, as humans, also have shared assumptions about human goals (that we can use).
Natural selection hasn't invented completely different goals for us all individually... So, we can use this to do Inverse Reinforcement Learning on human agents.

IRL cannot work for agents of unknown rationality, but we can use human's internal models as an assumption to do IRL on human agents.
We (and machines of the future) knew human internal models all along, and this should make it possible for us to work out what reward function(s) a given human is trying to maximize at a given time.
Simple, and yet pretty clever!

3.2. Topological Approaches for Unsupervised Learning.

Leland McInnes continued with a talk about ''Topological Approaches for Unsupervised Learning''.

Broadly speaking, topological data analysis has gained a lot of interest in some quarters, but, generally, has seen relatively little uptake in the broader machine learning community.
Even though, according to Mcinnes, ''topological approaches can be brought to bear upon unsupervised learning problems as diverse as dimension reduction, clustering, anomaly detection, word embedding, and metric learning''.
I.e. ''Through the lens and language of topology and category theory we can draw common threads through all these topics, pointing the way toward new approaches to these problems''...

That is - there are a variety of theories and techniques defined in topology, that are becoming (somewhat) popular within the machine-learning & deep-learning community. And, here, McInnes gave some quick hints and insights on how to get started with these ideas, and build them into your own Deep Learning pipeline.

Indeed, a good primer. For more, see the full talk here.

3.3. Parameter Servers Suck, All Hail Horovod.

Ruksi Laine, Valohai, gave a talk titled ''Parameter Servers Suck, All Hail Horovod ''.

More data is usually good, as model performance (i.e. the neural net performance) usually improves.
But more data also makes training (of the model/neural net) much slower.

So, how can we solve that problem? Throw in more GPU's?
Well, as Ruksi Laine brilliant explained, not so fast.

Deep Learning is innately sequential, and throwing in more GPU's will only help so much
(after a while, communication problems will slow progress etc).

In data-parallelization we run the same model over different datastreams, and then synchronize the weights, θ, in the model
(I.e. training different layers in a Deep Learning model in parallel on different GPUs would then be ''model-parallelism'').
In a very helpful way, Ruksi then went through a number of (new trends) ways to do distributed training, highlighting various bottlenecks in these methods. Super interesting, and very helpful to get this overview!

Having a parameter-server (that parallel processes should use) didn't appear to be a good approach...

Ruksi then went on to explain how he was using TensorFlow and Horovod (a distributed training framework for TensorFlow)

Horovod, a component of Michelangelo, is an open source distributed training framework for TensorFlow and its goal is to make distributed Deep Learning fast and easy to use via ring-allreduce and requires only a few lines of modification to user code.

Altogether, it was an excellent introduction on how to get started with these techniques.

3.4. Spot the villain - The Merlon Identity Index.

Dušan Fedorčák continued with a talk about helping banks to fight money laundering.

Dušan Fedorčák writes on his linkedin page:

Of course it's difficult. I like it that way.

And screening of potential bank customers does indeed seem to be a difficult task.
Fedorčák writes: ''Currently, we process more than 10 TB of news articles, and our main goal is to index the data for efficient search and risk identification''.

Clearly, it would be nice if you could just access a ''villain database'' in order to find the bad guys.

But how to build such an index?
Crawling the internet is fine, but most of the interesting information, that would allow you to associate names to good or bad behaviours, are probably not available for free on the internet. Data is not cheap.
Also a problem is that banks are pretty conservative, and will not accept any black box models. It should be explainable AI, where it is possible to explain why a potential customer is considered a risk.

Currently, the banks a) search for articles, b) read the articles and c) create a report.
So, it will probably be a good idea to start by doing something similar.

Fedorčák described (some of the) additional steps towards a solution (in the remaining part of the talk).
A demo of their solution is available at Merlon Intelligence:

We can help you level up your Financial Crime Compliance (FCC) capabilities by leveraging the very latest advances in Artificial Intelligence and Machine Learning to give you a single internal and external view of your customer's FCC risk that is continually updated.

Interesting!

3.5. Machine learning: Explainablity with anti-models.

Srivatsan Santhanam, talked about ''Explainablity with anti-models''.

We have (in ML):

• Models to Forecast.
• Models to Classify.
• Models to Cluster.

Here in this talk it was all about models to Verify/Infer/Explain what is going on.
Important, as explainability gives trust.

Interestingly, Santhanam described the deployment of a fraud detection (neural net) model that might have worked very well, but made everyone work slower, as all results had to be verified, so that people could trust what the model predicted, but didn't explain.

Other (ML) domains are equally eager to have explainable models.
E.g. ML models in healthcare are basicly useless, if healthcare professionals don't trust and understand what the (nn) models are predicting.

So, indeed, explaining complex machine learning models is a hot research topic.
And Santhanam clearly had some good insights on how to proceed.

Still, it is all work in progress, as I understood it... Something which will be refined in the future...
For now, I guess we can think of ''anti-models'', as ''unit-test'', light-weight models that users can play around with. Models that can help build trust in the end-users mind.
And we can then try to use some of the ideas from this talk as a starting point for our own experiments into building (more) explainable models.

Clearly, an important area, with much more to come in the coming years.

3.6. Solving the Text Labeling challenge with EnsembleLDA and Active Learning.

Alexander Loosley, talked about the ''Text Labeling Challenge'' in (many) text classification problems.

So, how do you get started when your boss drops a huge corpus of customer feedbacks, and wants to have labels on them, so they can be addressed in a more efficient manner?

One problem could be that it isn't so simple, at all, to find relevant topics in the text corpus (to begin with).
Loosley suggested that we look at EnsembleLDA (see LDA, compared to k-means, here).

I.e.

• Train an ensemble of topic models (Multiple learning algorithms to obtain better predictive performance than could be obtained from any of the constituent learning algorithms alone. Think e.g. random forest).
• Cluster the topics (based on ''distance'' between found topics). And, keep ''cores'' of clusters, ''reliable'' topics, and throw away everything else.

Code for such things (EnsembleLDA) can be found here:

Gensim is a Python library for topic modelling, document indexing and similarity retrieval with large corpora. Target audience is the natural language processing (NLP) and information retrieval (IR) community.

See:

• EnsembleLDA on Github.
• Misc. NLP material (RaRe, practical material for Machine Learning and Natural Language Processing).

At the end of the talk, a few hints were given for those wanting to dig more into the efficient labelling with active learning problems.
There weren't many details in the talk itself, but, on his blog, Petr Lorenc has written down a few keywords it might be useful to explore further (for those who are interested in digging more into this):

Do: Semantic Text Embeddings (eLMo,Bert) -> Do: Text similarity (Word Mover Distance) -> Do: Document representation plotted in 2D (tSNE) -> Use: Nearest neighbor models (kNN).

3.7. The Labels are Out There.

Loted Peled, talked about ''Labels''. I.e. we need labels in order to learn.

Certainly, supervised machine learning algorithms need labels in order to learn. So, how do we get these labels?

If we can find labelled datasets online - fine. But if not, Peled suggested that we should ''look for data sources that aren't necessarily labelled as datasets''. In her own work she had used online TV manuscripts to help with punctuation, when writing down spoken language (Doing speech recognition).
She also suggested that we should look for expert labeling via platforms such a fiverr.

Indeed, again, a very helpful talk.

More conference impressions... E.g. see ...

4. Impressions from Sunday, February 24th.

4.1. Solving the 3 main theoretical puzzles of Deep Learning.

Tomaso Poggio, MIT, talked about ''Solving the 3 main theoretical puzzles of Deep Learning''.

Poggio started the talk by stating that:

At the center for Brains and Machines:
We aim to make progress in understanding intelligence. We believe that the science of intelligence will enable better engineering of intelligence.

And posed the interesting question ''Why is there so little in terms of a theory explaining why deep networks work so well''.

Basicly, in deep learning, we are trying to find the weights that minimizes an error-function
(I.e. we are moving around in a fitness landscape for weights looking for a minima).
One could then ask why deep networks are better than shallow networks...?

I.e. as data dimensionality increases, and the sparsity of the data increases, this makes it harder to ascertain patterns. But why is it that deep learnings inherent nature (with many layers, where activation functions delivers results from one layer to another, with new activation functions etc., in a compositional way) gives better results (See: Curse of Dimensionality). Clearly, the layer structure of the deep networks, inspired by neuroscience, helps us preprocess, encode, problems, and helps us get better results (Local computations that can be combined to more complex functions are helpful).
But, in my understanding, here Poggio reminds us that we could be more precise (formal) in our descriptions of what is actually going on here.

The next question one could ask oneself (concerning deep learning networks), is why don't the gradient descent methods (for finding the weights of the networks) get stuck in local minima?
Interestingly, we were shown how it is possible to show that stochastic gradient descent methods finds (with high probability) the global minimum. I.e. which then assures us that the algorithms that we use actually work ... great...
.
So, optimization works with a lot of parameters, and the next question then becomes ''how to avoid overfitting''. With more parameters than data, the error in training goes down to zero. Still, it is usually possible to get decent results on the testdata. As we are selecting one solution out of an infinite amount of possible solutions.
But, understanding the exact relationship between parameters and datapoints (for good solutions) is clearly not trivial.

The talk ended with some thoughts about the evolution of computer science:

Evolution of Computer Science.

• There were programmers.
• There are now labelers.
• There may be schools for bots.

Now, we are in the ''Machine Learning Phase'', where Deep Learning Networks learns based on labelled data.
In the next phase we want to learn from (based on) just one or two training examples, i.e. in the same way as children learn...
But many more interesting questions are, of course, coming up. Questions like: a) How are symbols represented in networks? b) How can networks be trained in more biological plausible ways (than Backpropagation, Gradient Descent). c) Shouldn't we include time, how inputs come in time, in our models etc. etc. ...

Indeed, interesting times ahead!

4.2. How to win a Kaggle competition, and get familiar with machine learning?

Marcin Szeliga, talked about ''How to win a Kaggle competition, and get familiar with Machine Learning in the process''.
Or, simpler, ''Machine Learning best practices''...

According to Marcin Szeliga, clearly, good models are important in Machine Learning.
But it is also imprtant that we have good tools... Indeed, good tools are a prerequisite to succes in Machine Learning, when we are talking about things like:

• Performance
• Scalability
• Security
• Reliability
• Availability
• Ease of use

So, get some good tools...

Next, Szeliga argued (in the talk) that we should use some old and well-tested methodologies for dealing with ML problems (Some going back to 1997, and earlier. Back in the days, when Machine Learning was called Data Mining...).
He ended up recommending going through these 6 steps:

1. Problem (Business) Understanding (What is the problem all about).
2. Data Understanding (Data visualization). When it doubt, visualize, look for outliers etc.
3. Data Preparation. Move on to do feature engineering. Remove outliers, Impurity in data etc. Then, transform all (non-numerical) data to numerical data formats, as learning algorithms need this (Btw. - According to Szeliga, 20 years ago this step used to take 60 % of a projects time, now it is more like 85 % of the time). Then: Simplify, remove less important features...etc.
4. Modeling. Then we tune the hyper-parameters of the model.
5. Evaluation. Divide the data into subsets. Use some parts for training, others for tests (of the models). Choose the simplest model among the best (most accurate) models.
6. Deployment (He ended up being in the top 2% of the Kaggle Titanic competition).

So, follow these easy steps. Get something up and running, and then you are almost done.
Indeed, just do it...

More about (getting started on) practical ML projects here.

The Rudolfinum, Prague - Old world charm.

4.3. AutoML in Predictive Modeling.

Pavel Kordík, Unico.ai & Recombee, talked about ''Automated Machine Learning Algorithms''.

Lately, ''Automated Machine Learning'' has gained momentum (since 2018), with several open source projects released from top AI labs. Pavel Kordík told us about some of these projects.

E.g.

PocketFlow is an open-source framework for compressing and accelerating deep learning models with minimal human effort ...
Developers only need to specify the desired compression and/or acceleration ratios, and then PocketFlow will automatically choose proper hyper-parameters to generate a highly efficient compressed model for deployment.

AutoKeras is also pretty interesting:

The ultimate goal of AutoML is to provide easily accessible deep learning tools to domain experts with limited data science or machine learning background. Auto-Keras provides functions to automatically search for architecture and hyperparameters of deep learning models.

Pretty easy to getting started with... (In AutoML you don't have any parameters, so everything is done automatically ...)

import autokeras as ak

clf = ak.ImageClassifier()
clf.fit(x_train, y_train)
results = clf.predict(x_test)

Google's version is called AdaNet (build on top of TensorFlow):

AdaNet (Adaptive Structural Learning of Artificial Neural Networks) builds on recent AutoML efforts to be fast and flexible while providing learning guarantees.

See more about AdaNet here.

In all cases, AutoML automate model selection and hyperparameter tuning, based on performance and simplicity.

Super exciting stuff!
With more from Pavel Kordík here.
See also: prg.ai.

4.4. ML for Recommender Systems.

Marc Romeyn from Spotify talked about ''Recommender systems''.

Lots of interesting details - E.g. ''why do people in Denmark listen to Swedish songs, but not the other way around'' ?
Well, maybe the Swedish songs are better? :)

Still, again an interesting talk, with a lot from the Spotify Labs. Super!

For more, see Jakub Bahyl's blog about the talk (at ML Prague 2019).

And remember to check out Marc Romeyn on Github.

4.5. Spelling Corrections for Web Searches.

Vladimir Kadlec from Seznam.cz talked about ''Spelling Corrections for Web Searches''.

Seznam is a Czech site, which has many different services, including a news section, TV, free email etc. And 10 million queries per day.

About 10 % of all queries are misspelled, so we need spelling correction methods for these queries:
First the query is normalized; Accents are removed, the query is put in lower case etc.
After that they look at candidates for what the user might have wanted to write. Based on words were the letters are close to what was typed (on the keyboard). See Levenshtein distances (And, the Levenshtein algorithm, where a distance between two words is the minimum number of single-character edits). Etc.
The queries are then ranked according to a ML algorithm, XGBoost.
From top candidates (for the query) they then select a query.

We are almost there, but they do add some extra magic ingredients, including more ML stuff... probably... before they actually hand over a result to their customers...

Take home message: There is apparently a lot of ML going on behind the scenes at Seznam. Cool!
Check out the site here.

4.6. ML powered Crime Detection.

Kateŕina Veselovská, ''Nlp Girl'' talked about ''ML powered Crime Detection''.

We were asked to imagine ourselves committing a perfect crime, a murder...
''When you kill someone, it might be an advantage, if it looks like that the person committed suicide...''.
Of course ...

''And that the murdered person has left behind a suicide note that looks kind of authentic...''
Of course ...

Then how should we, now, as police officers, then tell real suicide notes apart form fake ones?
Machine Learning, of course ...

Give your ML algorithm a lot of real suicide notes, and a lot of fake ones, with classifications, and then lets see if it can figure out whether the suicide note is fake or not...

It turns out that the authentic ones have a) longer sentences, b) specific vocabulary and c) neutral sentiment, whereas the fakes one have a) shorter sentences b) generic vocabulary and c) positive sentiment.

Indeed, it is not easy to do crime in the age of ML.

Everything we write makes it pretty easy to identify us.
We all write with a specific word and sentence length.
Our punctuation also tell a lot about who we are. As do dialects, and overall grammar.

Age and gender is usually easy to guess based on a text. Just as it easy to guess social and local origin of the author. The level of education. Etc.
So, ML can, of course, find such things in a text...
(or at least, will be able to in the near future).

Indeed, a super interesting talk about ML methods for automated unstructured content analysis. Super cool!

4.7. Artistic Applications and Artificial Intelligence.

Luba Elliott talked about ''AI and Art''.

Luba Elliot started by reminding us about Alexander Mordvintsev, Christopher Olah & Mike Tyka's Inceptionism, where Neural Networks create ''art'':

We train networks by simply showing them many examples of what we want them to learn, hoping they extract the essence of the matter at hand...
...
Turns out that Neural networks that were trained to discriminate between different kinds of images have information inside them needed to generate images too...

When the neural nets have been trained, they then

Start with an existing image and give it to our neural net.
We ask the network: ''Whatever you see there, We want more of it!''. I.e. we ask the network to enhance whatever it detected (By gradually tweaking the image towards what the neural net detected ... and making sure that the picture still follow some basic image rules, such as that neighboring pixels need to be correlated...).
...
This creates a feedback loop: if a cloud looks a little bit like a bird, the network will make it look more like a bird.
This in turn will make the network recognize the bird even more strongly on the next pass and so forth.
If we apply the algorithm iteratively on its own outputs and apply some zooming after each iteration, we get an endless stream of new impressions, exploring the set of things the network knows about.
We can even start this process from a random-noise image, so that the result becomes purely the result of the neural network.

See more here.

She then showed us Gene Kogans ''Mona Lisa'':

Along with work from Sofia Crespo.
Some underwater creatures looked a little scary, but not as scary as Mario Klingemanns Neural Glitches.
Klingemanns Neurographer twitter feed is also interesting...

Peter sees the computer.
''But the machine only creates what humans have taught it to,''
says Peter.
''So do you,'' says Mummy.

According to Elliot, Roman Lipsky ''has a dialoque with the algorithm''. Well, don't know about that, but it looked interesting!

Scott Kelly and Ben Polkinghorne's Signs of the Times didn't look all that advanced techically. Indeed, it looked more like a somewhat lighthearted happening about the digital world out in the real world? Still cool though!

Constant Dullaart's homepage makes you wonder what Van Goghs homepage would have looked like, had he lived in an era with homepages...Weird, for sure... Certainly, it takes a little getting used to, but I guess it is ''different'' enough to be interesting!

MegaPixels.cc is

an art and research project investigating the ethics, origins, and individual privacy implications of face recognition datasets created ''in the wild''.

by Adam Harvey, who also has some great ''data pools'' on his website ...
Awesome !

More at AI art online.

4.8. SketchPad AI.

In e.g. SketchPad the computer doesn't really help you draw something.
If you can't draw a horse - Well then you can't draw a horse ...

Enters CEAI, and other great AI startups...
Surely, you should only draw a little bit, and then the computer should be able to identify what you are up to - and try to complete the drawing?

Tried it at the conference, and it worked surprisingly well (sometimes...).
Very cool, nevertheless!

Indeed, it was an awesome day!

4.9. Meeting Pepper. Prague Airport, Sunday February 24th.

No ML conference would be complete without robots to deliver a wholesome atmosphere...

And sure enough, here in Prague, the ever friendly Pepper greeted people in the airport, and guided them towards their destinations.

Click on Pepper pictures above for my Youtube Pebber (in Prague) video.
And compare with my (CogSci 2014) Poppy Robot video.

5. Prague impressions.

Misc. post-conference impressions from Prague.

5.1. Mozart.

Mozart came to Prague on the 4th of October 1787 to help supervise the first performance of his opera Don Giovanni. The premiere was in the Estates Theatre on the 29 October 1787. The Prager Oberpostamtzeitung reported, that ''Connoisseurs and musicians say that Prague has never heard the like''.

The statue of the ghost outside of the theater represents the operas character ''Il Commendatore''.

Youtube Clip: Don Giovanni - Commendatore.

5.2. Dvořák.

Astronaut Neil Armstrong took a tape recording of Dvořák's Ninth Symphony, ''The New World'' along during the Apollo 11 mission.

Earlier, on Saturday, 4 January 1896, the Rudolfinum witnessed the inaugural concert for the Czech Philharmonic, when Dvořák, himself, ascended the stage to conduct the then already world-famous Symphony No. 9 ''From the New World''.

There is now a statue of Dvořák outside the Rudolfinum (The Rudolfinum is, btw., named after Crown Prince Rudolf, who sadly, was mentally unstable, and on January 30th 1889, was discovered dead along with his mistress, baroness Vetsera, as the result of an apparent murder - suicide).

Youtube Clip: From the New World.
Along with Smetana's The Moldau among the best known examples ever of music to come out of Bohemia, Moravia, and Silesia (Now the Czech Republic).

5.3. Memorial Plagues on house walls in Prague.

Spotted a number of interesting memorial plaques, as I walked around in Prague.

And, NB, my android phone did, of course, log my route, Saturday in Prague. Just in case I had any doubts, precisely, where I had spotted these memorial plaques.

Here in the salon of Mrs. Berta Fanta, Albert Einstein, professor at Prague University in 1911 to 1912, founder of the Theory of Relativity, Nobel Prize winner, played the violin, and met his friends, famous writers Max Brod and Franz Kafka.

Einstein would later move on to Berlin, where there is another memorial plaque (in the Humboldt Universität), See my CogSci 2013 report, here.

5.4. Franz Kafka in Prague.

Kafkas birth house in Prague.
See BBC's mini guide to Kafka's Prague for details.

Still standing, traces of the old jewish part of Prague that Kafka was a member of.

In Jewish folklore, a golem is a mystical being created entirely from mud or clay. Rabbi Judah Loew created the Prague Golem to protect the Jewish quarter of Prague and its citizens.

In Kafka's day, Prague's coffeehouses were the place to meet and talk, about it all...
Kafkas favourite place was the Cafe Louvre, established in 1902, and still standing.
And no visit to Prague is, of course, complete without a visit to the Louvre. For a coffee and some existential angst...?

Statue of Kafka by Jaroslav Róna in Prague:

The awesome kinetic Kafka head,
outside the Quadrio shopping centre in Prague, by David Cerný. For more details, see my youtube video of the kinetic head.

Kafka worked for the Insurance company "Worker's Accident Insurance Institute for the Kingdom of Bohemia".
The company building is still standing, and is now a hotel (Hotel Century).
Asked in the lobby, if this was indeed Kafkas old insurance company, which they immediately confirmed, just as they pointed to a statue of Kafka (in the lobby), in case I wasn't fully convinced...

Indeed, the very friendly receptionist was happy to show me the Kafka highlights in the building.
Amazingly, it turns out that you can actually rent Kafka's old office, and stay there as in any other hotel room.
Breathing in Kafka's work-life as you go to bed... surreal!
And btw. Kafka couldn't have been such a terrible insurance employee, as I imagined. My friendly receptionist/guide told me, standing in Kafka's office... that Kafka was actually promoted, while working in the insurance company, and got a bigger office...

6. Conclusion.

The end of a wunderbar conference. With many memorable talks.
Obviously, I'm already looking forward to my next visit to Prague!

More pictures from Prague, February 22 - 24, 2019.

Conference Venue.
Rudolfinum. Prague, Czech Republic.

Time to pack up & say goodbye.
And, perhaps, meet again next year in Prague...

The end of great conference !