11 years of Android: A brief history of time

Hello dear reader. My name is Ivan and I did 11 years of Android so you don’t have to!

Yes, this is a cautionary tale for any young mind foolish enough to venture into the dark depths of the tech industry. It is a personal story full of excitement and wonder, ups and downs, struggles and breakthroughs, trials and errors, exaggeration and oversimplifications for your entertainment. It is also filled with domain specific engineering jargon like: domain specificcontext leaksif (!isAdded()) returncrippling anxietycoroutine exception scavenger huntscreen orientation lock,

Each activity that you lock into portrait adds an additional week to your life expectancy

From what you’ve read so far, you might assume that I dislike Android. That is not the case. It would be silly to spend 33% of my entire and 100% of my adult life working on something I don’t like.
I do like Android, in fact. It made me how I am today. It allowed my financial freedom and mobility. It allowed me to work with amazingly brilliant people on awesome ideas. I’ve traveled, advocated, started, failed, learned, managed, led, created, connected thanks to Android.

But, we have to be real: Android development is not great!
It never has been. It may never be. Sure, it has gotten better lately, but it is not the prince that was promised. It is just a John Snow.

My bet is on Bitcoin now.

To walk you through my bold claim, I would match my subjective experience with Android to the historical periods of human development, starting from prehistory all the way to the present post truth era. Brace yourselves for a brief history of time:

The Prehistory

Yes, you read it right, prehistory. You may not be personally aware of this, but we are on the exponential curve of development heading straight towards singularity. So many things have happened in recent years, 11 years ago does seem like prehistory. Few hundred years ago, if your father was carpenter, chances were you would be one too (under the assumption that you managed to survive your childhood). Nowadays, if you choose one framework, chances are in 12 months you would need to refactor everything. *coughs* RxJava *coughs*. Except for Fragments of course. They are here to stay. Like the appendix in our body, Fragments are laying dormant, waiting for someone to restore the instance state or pop the back stack and wreck some havoc.

Little is know about prehistory. The tools we used were made from wood, they did not survive the elements, they decomposed over time. What we can recall is from the paintings on the walls of the deepest caves. The Activities were the deities, the God classes, containing all logic, leaking contexts left and right. Things that give modern developers a mini heart attack today were common back then. Manual network requests, manual image loading and caching, network on main thread, yes, network on main thread, dreadfully slow and unusable emulator, trackballs, hardware keys, jars, Eclipse IDE, svn, SOAP. Simple Object Acquisition Protocol. Oh my, what a misleading name. The person that made this up sure had the last laugh.

The only SOAP you should be aware of. Stay healthy, stay strong in these odd pandemic times!

Indeed a time of great hardship, but also a time of great freedom. No design guidelines, zero requirements for publishing on Android Market (yes, Google Play was called Android Market once), no libraries to tell you to initialize them in the Application class, no doze mode, no limits to background execution, no data to process, no stack traces to analyze. The path to greatness was straightforward, you just copy the first answer from Stack Overflow, you don’t really need to understand what attachToRoot parameter means during layout inflation.

The Ancient History

Ancient History, referred also as the Classical Period, named after the emergence of the classical Android problem: fragmentation. Every person in the world with a stable internet connection downloaded the AOSP code and made their own version of Android. You want Hello Kitty icons? You got it! HTC UI on a ZTE device? Why not! An ad blocker with root privileges? Yes sir! Oh, the Bluetooth is not working, the camera only captures portrait and there is a high possibility for bricking? No worries, the cyan Android is worth all that.

This made me the most popular person in my crew after the people that owned iPhones

Tiny civilizations formed around these early attempts of consolidation, each with their own similar, yet distinct culture. The platform was maturing, the Contexts in the AsyncTasks were being wrapped in WeakReferences. JSON was the language most servers could speak. Git became the one version control system to rule them all. Steve Jobs unleashed an oversized phone (that would later on be referred as tablet) which made development needlessly more complex. Whispers traveled from the far East about an exotic IDE and a new build system that will make development a breeze. No more jars, no more ants.

Language evolved as well. People started talking about and partially understood the previously unheard topics of ViewHolders, Fragments, Tablets, Holo Themes, SherlockActionBars, Hamburger Menus. There are written evidence of Android apps that managed to look good while NOT being an exact copy of the iOS one, a groundbreaking progress made by early design pioneers and courageous explorers.

While there were sparks of civilizations here and there and the future looked bright (ex. Instagram just got ported to Android), keep in mind these were still barbaric times, a wild free for all with little structure, excessive wake locks, permanent background processes and massive battery drainages.

The Dark Ages

At this point of history, Android became the weapon of choice for the masses, offering a wide variety of mostly functioning low/mid end devices. Blackberry was moved to museums next to where the dinosaurs were kept and while still early for a full on flagship battle with iOS, the ecosystem provided everything you would need from a smart phone for a price that would not require your parents to get a short term, high interest loan for you to look cool. It may sound silly today, but these were times when iOS vs. Android rivalry was taken quite seriously.

The worthy knights had to pledge their alliance, worship and defend their consumer preference honour vigorously.

According a scientific census that I made up for the purpose of this article, the dark ages officially began when the first production app hit the 64K methods in a single DEX file limit. This signalled that the limits of the intended platform usage have been reached and now fragmentation is running amok. This, together with few other cataclysmic events like Samsung consolidating a large chunk of the market with the worst phone in the history of phones – S3, brought upon the dark ages.

In the fist company that I worked in, we had a whole team dedicated to handling S3 specific bugs.

Google did their best to fix the leakages, address the platform bugs, force OEMs to push OTA updates of their OS (if you don’t know what these abbreviations mean, just read forward and save yourself a google search) and provide useful APIs and guidelines to developers, but to avail. The manufacturers were slow to react, Views were filled with if-else branches for handling device/platform specific bugs and 3rd party libraries were using AlarmManagers to upload entire phonebooks, call logs and messages to their servers while you were sleeping.

Development felt more like trial and error, patch and ship than a structured approach. It still took three Hail Mary’s and a process death to start up the emulator (hello Genymotion), build speeds slowly deteriorated and up until this day they never fully recovered. It became a common practice development teams to have dummy projects that compile within reasonable times, develop features and especially UI there, then just copy paste them to the main project. It did also not help that Google was pushing for the worst API ever created for Android – Loaders to compete with already terrible ones like AsyncTask, ContentProvider, ContentResolver, ListView, RecyclerView (View in general for that manner), Camera, Bluetooth, Resources, Styling, Theming, Keyboard,…

The Renaissance

Regardless of how dark it gets, human ingenuity will never let darkness prevail. The Renaissance, the great awakening, the age of enlightenment started when Google figured out that you don’t really need to wait for device manufactures to push fixes updates to their users, they can delegate that responsibility to the developers. Thus, the support libraries were born, a massive overhead of code and complexity that you need to include and maintain in your project to patch an OS that cannot update fast enough to compete with Apple.

This forward thinking started rolling the ball forward and brought even forwarder thinking. Just like the great renaissance Renaissance painters that drew inspiration from Ancient Greek culture, Android developers wondered what can be learned by studying ancient cultures. Was there any wisdom hidden in backend or web development? Did they know something we had forgotten?

What is this cryptic language trying to tell us?

It turned out there was! For example, you can actually organize your code, not just stuff everything inside an Activity or a Fragment. By doing so, you can split the presentation from the domain layer and even test them separately.
I did not believe it when I first heard it. Writing your code in a way that is testable and reusable seemed wasteful. What was next, writing your tests first, then writing your code? Heresy! I wanted none of that.

But, just like in human history, science prevailed. Uncle Bob figured it out. It turned out that the MV(C/P/VM/I) is a good way to organize your code and all the boilerplate is just worth it. Sure you would copy and paste the same thing in dozens of different files for a while, but it is a matter of time until someone writes an annotation processor or a command line tool that will generate this for you, confusing the Gradle build system and setting you on a google search quest that will ultimately lead to deleting the build folder, invalidating the cache and restarting the Android Studio.

Support libraries grew to become behemoths (remember when just including the GSM megalith pushed your app over the 64K limit), but also enabled creating better user experiences. Android was not the underdog anymore. An opportunity for taking the lead arose. For the first time in history Android became cool.

Of course, some forward thinking went too far. Reactive programming became a fad, but ultimately caused more problems that it solved. Same goes for any ORM libraries that tried to abstract out the limits of SQLite (hello Realm). Google also got excited and went ahead of themselves with Ahead Of Time compiling, Jack, Instant Run,… All these ideas were awesome in theory and received massive highlights during Google I/O events, but ultimately failed to deliver on the promise of making Android development great.

The Industrial Revolution

Once people find something that is working, they are amazing in scaling it up, iterating over it, finding more and more ways to squeeze every last drop, consolidating the market, removing the noise, silencing the competition. Enter the machine dear developer, rejoice! Gone are the days of you being stuck with a bug from a dependency because the repo maintainer went on a vacation and never came back. Gone are the days where you would have to rely on Jake Wharton‘s pet projects to find views by ids in a fancy manner. The capitalists took over. The old gods are dead. You will heed to what Google “recommends”.

And if you don’t like it you can always use Huawei‘s services

Development was moved to the assembly line. It was not longer art and experimentation. It was no longer “we do what our introvert tech lead genius tells us to”. We now know what works: Kotlin + MVVM + Coroutines + the full Google suite of support libraries and services. You inject stuff now. Inversion of control or something. You don’t need backend even, you have Firebase that cost practically nothing under the assumption that you excessively duplicate and nest your data to avoid multiple queries. Everything is documented and everything works as intended except when it doesn’t.

You don’t even need to think, the Android Studio new project wizard generates all the boilerplate for you, sets up the navigation, the annotation processors that never affect your build speeds, decides what flavours to be built on your CI/CD build machines, it tracks every user action and reports their feedback in a Slack channel that the PO/tech/team lead needs to monitor to demonstrate to management that all that talk that branding and marketing came up with about caring for our users is actually true.

Sure, you still need to write Adapters and bind ViewHolders for your RecyclerViews like is 2010, you still have to guess if the keyboard is up or not, you still need to google on how Service can communicate with an Activity, you are still uncertain what library update will break the incremental compilation, forcing you to think that at this point the only thing that may reduce build speeds is Apple Silicon.

But, that is all fine. At least now you have social login out of the box, mature design system – Material with solid principles and extensive guidelines that everyone knows, except for some reason, your designer. That is not all, you also have streaming operations for lists and extension functions you can use to puzzle the junior developers as you chain your way to the most elegant piece of code you have ever written.

        viewModelScope.launch {
            firstFlow
                .zip(secondFlow) { first, second ->
                    first.takeIf { it.contains(second) }
                }
                .mapNotNull { it }
                .distinctUntilChanged()
                .collect { Log.i("TAG", it) }
        }

The Modern Times

People in the early 20th century envisioned the future as this bright, modern place with flying cars and Jetpack. No flying cars yet, but at least we have Jetpack now to fly us places. Showing late to the party, Android finally embraces declarative UI with Compose, hopefully just in time to prevent some young kids venturing into Flutter or god forbid React Native with their false promises of covering multiple mobile, web and desktop platforms with a single codebase, hot reloads, a simpler way to deal with UI, navigation, multithreading, theming, styling, animations,…

But what about performance fellow kids?

Updates nowadays are more incremental polishes than meaningful changes. Except for the top notch, that was a game changer, it triggered a movement of millions of developers spending at least few hours adapting their layouts to account for the brilliant stunt that Tim Cook and Jony Ive pulled to break the $1000 mark for a phone. Can wait for the $2000 mark for the foldable displays.

Edges alternate between square and rounded every season, because groundbreaking design progress happens once a decade or so. In between you go from AlertDialogs to BottomSheets, from Rose Pink to Crimson Blue. After more than a decade of maturing, UI or end-to-end flow testing is still not viable, the amount of effort that takes to setup and maintain is really not worth it, business wise is easier and/or more cost efficient to outsource and let people with names you struggle to pronounce to click through hundreds of test cases. You can use your precious developers time to round off that Toolbar, so your app does not look exactly like million others in existence.

The European Union with their concern for user’s privacy in collaboration with hackers interested in obtaining that privacy, shrank the limits of what you can you as a developer. Permission and consents pop up all the time, users promptly deny or ignore them, depriving companies from the much needed data, resulting in even a stronger push to get it by any means necessary.

And this is what modern times development is, tricking the users into giving up their data. It is less about solving a problem or providing a service and more about figuring out an ad placement that will yield +0.6% of revenue through a masterfully crafted A/B test. It is about integrating tracking libraries to segment your users and serve them personalized content based on a model that your data team trained and your backend delivered via Server Driven UI.

Modern times come with modern fancy tools, you can now monitor network, memory, CPU, object (de)allocation, memory leaks, query local databases, peek into the layout hierarchy, drag-and-drop constraints,… Still, most of the time you will not spend using them, but figuring out why they need plugging and unplugging the USB cable in order to work or something completely unrelated like why the horizontally scrolling nested RecyclerView within a vertically scrolling parent one does not retain the state upon orientation change.

The Build Analyzer is here to ensure you that there is nothing you can do to improve your build speeds until you remove all that legacy Java.

The Future

If we 2012 and the Maya ever thought us anything, is that no one can really predict the future with great confidence.

If I was able to do so, I would not have written this blog, but race Elon to Mars instead.

Android has matured so much, that I can imagine being somewhat difficult for new developers to catch up. Activity/Fragment/Service lifecycles, Adapters and ViewHolders, FragmentManagers, ViewModels, LiveData, the complex topic of multithreading with the added spice of coroutines and of course, Flows can be quite overwhelming if you are just starting.

In my free time I play around with ReactNative and Flutter and I will now say something that myself from few years ago would never dare to say it: It is so much better! Not the best, these platforms still have some maturing to do, but much better.

Maybe even Kotlin Multiplatform one day, TypeScript and Dart are just so last century.

And I can recommend these frameworks (with a preference towards Flutter, but this is personal) to people who would want to get into mobile development. It just makes more sense. Why limit yourself to just Android by learning Native when you can pretty much do the whole frontend with a single framework. Mobile, web, desktop and whatnot. Include few Firebase dependencies and you are full-stack now.

And as more and more beginners are starting with multi-platform, as more and more tech leads are pushing the idea to management that a full refactor is needed to sustain the business, as more and more complex use cases emerge, these frameworks will mature and not be just a great tool for pushing out MVPs to get series A round of funding and give up significant part of your start-up ownership to people that have more money that they can ever use.

Fear not, Android Native will not die out. There will be plenty of work for Native Android developers in the coming years. It still needs to exist in order to support the multi-platform, there are still specific use cases where multi-platform will not work yet (ex. IoT) and there are huge companies around the world that have invested in Native so much, that it will not be easy to switch to something else.

The OS will still exist, there are individuals and huge teams at Google, Android, Huawei that continuously contribute (albeit, on their own forks of the AOSP) and evolve the system to fit in many other places, not just your smart phone.

I am still nostalgic about the timeline in the parallel universe where Google Glass became a thing.

And there you have it, a brief history Android. A small glimpse of the evolution of a framework through my subjective experience.
If there is anything to be learned from all of this is that change is the only constant. Things change at such a fast pace that is practically impossible to try, catch (hehe) and learn them all. And that is perfectly fine. There is a joy of missing out, not just fear.

The wisdom I want to leave you with is not of a technological nature, but of a human one. Take it easy. Do not value yourself by how much frameworks you are proficient in, trends you follow, buzz words you know, libraries in alpha you have in your pet project, lines of code you push to your side hustle after work hours. Do not gamify your profession jumping from one achievement to another, comparing yourself with others under the pressure of being constantly productive.

Instead, value collaboration over competition. Find the value you are providing to your team, your company, your product and ultimately, the end user. The greatest challenges our industry faces are not from a technological nature, communication is the greatest friction point. How do I get in touch with my inner self, so I understand how I feel, begin to understand others, and develop the necessary skills to communicate this efficiently with the people I work with. Your career progress is much less dependant on what you know and can do, but on how well can you learn and incorporate that in a collaborative environment.

True progress happens in teams and organizations that have understanding for personal preferences, practice open, honest, non-violent communication and nurture an environment that encourages learning through inclusion, empowerment, exploration and experimentation. These are the key words every tech person should be learning. Android, iOS, Fuchsia, Native, Multi-platform are just tools, they do not manifest ideas by themselves. We do, and we do so by working together.

Ivan is a yogi mindful leader, principal software engineer, a multi-platform enthusiast with a passion to bring mindfulness into the tech world. A passion so great that he decided to quit his day job and start a company together with his wife that does exactly that.
Ivan also gives talks on how to become a better human by making small tweaks and incremental adjustments to your daily life, the agile way. His next talk is titled 27 tips and tricks for navigating the pandemics (or any other hardship) and if you are reading this before 18 May 2021, you should definitely register by following the link. Thanks for reading fellow human!

Ivan Dimoski
Freelancer and member of DevBay community


Time series classification – an overview

Time series classification (Fig. 1) is a relative late bloomer in the field of machine learning. However, in recent years there has been a real surge of data, algorithms and open source code that can be used to tackle problems in new and better ways.

This blog post aims to give a basic overview of the field. My aim is that it should be possible to follow for the interested reader even without a background in machine learning. Hopefully this will give you readers an overview and further avenues for explorations.

Fig. 1. Time series classification example. Samples from the widely used synthetic Cylinder-Bell-Funnel (CBF) benchmark dataset [1]. Each time series belongs to one of three classes: cylinders, bells and funnels. Adapted from [2] with permission from Patrick Schäfer.

Examples of time series and classification problems

A time series is just one (uni-dimensional) or several (multi-dimensional) temporally ordered sequences of numeric values. They are ubiquitous since anything numeric that you measure over time or in a sequence is a time series. For example:

  • The temperature in Stockholm each day during 2020 (a uni-dimensional time series).
  • Sales records
  • Video (a multi-dimensional time series where each image corresponds to a time point. Note that here the data is also ordered in the pixel dimensions).
  • Audio recordings.
  • Internet-of-things and other sensor data.
  • ECG/EEG and other medical signals.
  • Transport data (e.g. when studying road congestion).
  • The growth curve of a child (here the time points are not equally spaced, so this puts special demands on the algorithms).

Other types of data can also be viewed as/transformed into time series, such as written text, which is basically a time series but where the entities are not numeric. The beauty of this is that it lends the possibility to analyze time series using language models and to analyze language using time series models.

Time series classification problems are everywhere, so it is hard to know where to start, but the following are some random examples of making classifications from time series data:

  • Classify of whom a voice recording is
  • Classify an ECG as normal or give the type of abnormality.
  • Surveillance: From a video, capture the path of an individual, then classify what he/she is doing.
  • Internet-of-things: classify whether a kitchen device is malfunctioning.
  • Speech recognition
  • Classification of brain imaging or genetic expression data

You can imagine yourself that the applications of good algorithms are essentially without limit.

What is time series classification with machine learning?

In essence, time series classification is a type of supervised machine learning problem. Supervised problems have the following procedure: You get a set of time series, each with a class label. You typically divide the time series into three groups, the training data, the validation data and the test data. You train a number of algorithms/models on time series in the training data, observe which algorithm performs the best on the validation data and choose that one. Finally, you use the test data to determine the performance of the chosen algorithm.

The difference to many classification problems in machine learning is that the data is ordered along the time dimension, and as such a good algorithm would need to exploit this property of the data.

One should not confuse time series classification with forecasting. The purpose is different and hence the algorithms are, too. Forecasting aims to predict the next future values, and as such often relies more heavily on the end of a time series. It doesn’t need to compare different time series with each other. Rather it needs to find recurring patterns in data that are predictive of the (immediate) future. Classification on the other hand needs to find patterns in the data that are different between different classes in order to determine the class of the time series at hand.

So what are some of the algorithms used for time series classification?

Dynamic time warping, a benchmark algorithm

For at least a decade, a technique called dynamic time warping (DTW) combined with 1-nearest neighbor (1-NN) has been a benchmark for other time series classification algorithms to beat [3]. 1-nearest neighbor (Fig. 2) is a simple technique where you have a training set of time series [4]. You then classify a new incoming time series by finding the time series in the training data that is most similar, and assign the new time series to the same class as that one. You can also have k-nearest neighbors, where you find the k most similar time series and choose the most common class amongst those.

Fig. 2. 1-nearest neighbor. Based on a dataset of two classes (blue and yellow), a new datapoint (circle) is to be classified. By finding the nearest datapoint, the new datapoint can be classified as belonging to the yellow class.

DTW [5] is used to calculate the distance between two time series (Fig. 3); without a distance 1-NN cannot determine which time series is nearest. The most naive way would be to just take the distance between each point in the time series. However, it is not necessarily clear which points should be compared to which in the two time series. For instance, if two identical time series are just shifted slightly, then this would result in a big distance. DTW solves this by pairing up the different time points by drawing lines between them in such a way that each time point in a series must be connected to a time point in the other series, and two lines must never cross (Fig. 3). The distance is the sum of the difference between the paired time points. DTW chooses the grey lines in a way that minimizes this distance.

Fig 3. DTW. Points in two time series are connected in a way to minimize the distance between two time series (the combined length of all the grey lines). As can be seen from the gray lines being tilted, the latter half of the blue time series is shifted to the right of the latter part of the red time series. DTW is able to connect the corresponding points of the two series to calculate the minimum distance. Adapted from [6] with permission from Brian Iwana.

 

The benefits of a machine learning model

DTW+1-NN doesn’t build a machine learning model of what time series for a different class should look like. What I mean is, there is no model holding an internal representation of the world, no knowledge stored in any model weights. Instead it reaches its conclusions simply following a step-by-step procedure where it compares new time series to series in the training data. The benefit is that no model training is needed. But there are several drawbacks.

Firstly, classification takes a lot of time. For each new time series one must calculate the distance to all the time series in the training data. DTW is in itself an optimization problem, which means that the distance calculation is fairly slow, and for a big data set consisting of millions of data time series it must be repeated over and over again. Hence, it doesn’t scale beyond fairly small problems.

Secondly, a machine learning model is a way to capture our knowledge of data in a parsimonious way. For a real-life analogy, having a model is akin to having an expert at hand. The expert can just look at your new time series and say which class it belongs to. Training a human expert takes a long time, but once that is done, classification is fast. On the other hand, DTW+1-NN is like working with a non-expert given a recipe for how to go through the data and determine the class.

Having once captured your knowledge into a machine learning model means that you can employ this knowledge in many situations that are similar but not identical, a technique called transfer learning. For example, consider a model that has looked at many bird songs in order to separate between nightingales and canaries. In order to do so, it has had to learn many features of bird songs in general. If we now wanted to separate between a robin and a sparrow, we could use our previous model and just finetune it to hear the difference between these two species. This would be much easier than training a new model, because the old model has a good ear for bird songs in general already, so much less training data would be needed.

You might think that transfer learning would only be useful in obviously similar cases such as the above. Think again. Image classification has been a resounding success due to the availability of huge benchmark datasets of millions of images coupled with competitions that have driven the development of algorithms to beat previous records on those datasets [789]. Researchers have for some time now transformed one-dimensional time series into two-dimensional images using techniques such as gramian angular fields, and been able to classify these time series images successfully [10]. Think about how wild this is! The network has been studying photos of cars and airplanes and somehow learnt something useful for classifying bird songs. When I first heard that learning can transfer that far I was just blown away. The reason why this works is that there is some similarity on some level between the time series images and real images on which the network was first trained.

The above example is interesting from another aspect as well. We have a 2-D image model which is sub-optimal compared to a tailor-made 1-D time series model. However, since the image model has had much more training data, it performs better. It is the model-data combo that determines the performance, not the model in isolation.

Recent developments / Other approaches

Data availability

For time series, there exists a repository of datasets called UCR/UEA [111] (from Univerisity of California, Riverside and University of East Anglia). It displays a wide variety of time series and all algorithms are benchmarked against it. In this way it is immensely useful in determining the relative performance of various algorithms. Compared to the corresponding benchmark repositories for images, text, etc, it suffers from being much smaller, having just on the order of 100 datasets, and with the biggest containing less than 10000 samples. For text, for example, you can download the whole of wikipedia very quickly. This comparative lack of data is hampering the development of the field. However, the repository is growing and being added upon every few years (most recently in 2018), which benefits research.

Perhaps due to relative lack of data, it has been difficult for new algorithms to beat DTW+1-NN, since more complex models would need more data to train all the parameters on. Below I describe some of the approaches that have performed better than DTW+1-NN.

Shapelets

One technique is to use shapelets [12]. A shapelet is a short time series that has been chosen because it is very indicative of a class (Fig. 4). For example, it could be a particular electrical pattern that is only emitted by microwave ovens and as such can be used when classifying home appliances. Shapelets are different from DTW because they focus on only parts of the time series. When you pass a time series through a shapelet algorithm the output is the minimum distance between the shapelet and all subsequences of the time series (this is called a shapelet transform). This data can then be used to train a classification algorithm, e.g. 1-NN or a machine learning model. In [13] an ensemble (see explanation below) of classifiers based on shapelet-transformed data was second best on the UCR/UEA dataset of all tested algorithms.

Fig. 4. Shapelets are small pieces of time series associated with a class. The distance between the time series (blue) and the shapelet (red) is the shortest distance between the shapelet and any part of the time series (here indicated by the dots). Taken from [14] with permission from Alexandra Amidon.

The main difficulty when using shapelets is that it can be hard to know which shapelets to use. One possibility is to manually craft a set of shapelets, but this can be quite difficult and time consuming. A better solution, if possible, is to automatically select shapelets. The way to do this best is an active field of research to a problem that is quite tricky due to the enormous amount of possible shapelets (see [15], section VI, and [13] for a good overview).

Another problem is really bad time complexity (how the execution time scales with the data): O(n2l4), where n is the number of training samples and l is the shapelet length [13]. This means that doubling the training data leads to a 4-fold slowdown and doubling the shapelet length leads to a 16-fold slowdown.

Model ensembles

Another approach has been to create model ensembles. “Collective of Transformation-Based Ensembles” (COTE) [16] from 2015 and its development HIVE-COTE from 2016 [17] are ensembles (collections, see below) consisting of many different models. COTE managed to increase the accuracy with 8% on the UCR/UEA datasets vs DTW+1-NN [13], which was the best of all algorithms tested. HIVE-COTE then significantly outperformed COTE and is the best algorithm to date that is not a neural network.

An ensemble is a collection of models each with its own classification, from which you pick the most common one. This works well if models have good enough accuracy and the errors that each model makes is different. For example, having three uncorrelated models with 90% accuracy each, for a given time series there is a 0.1*0.9*0.9 = 0.081 chance of one model giving the wrong answer, a 0.1*0.1*0.9 = 0.009 chance of two models giving the wrong answer and a 0.1*0.1*0.1 = 0.001 chance of three models giving the wrong answer. Taking the majority vote, the ensemble would give the wrong answer only in the latter two cases, representing 0.009 + 0.001 = 0.01 = 1%. Thus we have increased the accuracy of the ensemble to 99% vs the 90% of the single model.

An important factor for an ensemble is the degree to which the algorithms are uncorrelated (produce errors in different places). If all algorithms are totally correlated, they will produce the same results and so the ensemble will have the same accuracy as the individual algorithms. The authors of COTE noted that several of the 35 employed algorithms were minor variations of the same team (e.g. 50% worked in on auto-correlation and spectrum features). HIVE-COTE tried to address this by creating a hierarchy whereby all the algorithms were first grouped by type (DTW-based, dictionary-based, etc). Each major group first produced a joint classification and that was then entered into the ensemble classification. In this way they were able to avoid that the ensemble was dominated by one group of correlated algorithms, with the result of improved performance.

The biggest problem with COTE and HIVE-COTE is that several of the models rely on the shapelet transform and hence suffer from very bad time complexity, meaning that they take ages to run on a big dataset [13]. For this reason, the community has looked for faster alternatives. Another ensemble, TS-CHIEF, although not having higher accuracy than HIVE-COTE, is one such choice [18].

With so many models it is also very difficult to understand why the models classify the way they do. Interpretability is tremendously important. For example, it is not enough for a physician to say that you belong to the class of people for which there are little chances of survival in case of contracting corona virus and therefore we will not give you treatment.

Dictionary-based approaches

A third popular collection of techniques is called dictionary-based [13]. Basically, a dictionary is a list of words where you look up a word to retrieve something, e.g. an explanation of the word or the number of occurrences of the word in a text. So for instance one could classify a text as belonging to a type of field (e.g. sports vs law) by comparing the occurrences of different words.

But how does this apply to time series classification? Well, there are several ways to construct dictionaries for time series. One could use the shapelet approach above [1215] to count the number of times in a time series where there is a match between the shapelet and the time series. In such a dictionary the shapelet takes the place of the word and what you retrieve is the number of matches of the shapelet for the time series. An algorithm called ROCKET [19] is currently among the top-performing time series algorithms and does something similar. It has a large number of random kernels (very similar to shapelets) and they calculate for each kernel the percent matches. A simple classifier (logistic or ridge regression) is trained on the dictionary.

Another is the Bag-of-Patterns (BOP) algorithm [20]. It actually underperforms against DTW+1-NN [13], but I include it here because it illustrates the approach well. Like the Bag-of-Words algorithm used for text data, this algorithm counts the number of times a word occurs, only this time the words are created from numbers using a technique called Symbolic Aggregate approXimation (SAX) [21]. A bit oversimplified, you first split the time series into a number of larger time segments and normalize. Next you define intervals of the signal (y) value and assign a letter to each (e.g. A: <-0.3, B: -0.3 to 0.3, C: >0.3). You finally cut each segment into small pieces, calculate the mean for each piece and translate it to one of the letters (A, B, C in this case). In this way each segment consists of a string of letters, e.g. ABBC. You can then summarize the number of occurrences of each string into a dictionary, which can be used for classification.

Other dictionary approaches have been more successful, in particular an algorithm called BOSS [2]. It works similarly to BOP, but instead of translating the normalized original signal into words it works by Fourier transforming the signal into frequency space, low-pass filtering it and translating it into words using two algorithms called SFA and MCB. This has the effect of reducing noise. Furthermore, it creates an ensemble of different models that vary by the amount of low-pass filtering and the word length. In a comparison between many different algorithms on datasets from UCR/UEA, the BOSS algorithm was among the best, and is much faster on big datasets than COTE (complexity: O(n2l2)) [13].

One of the main reasons why dictionary based approaches work well is that the translation into words make them robust to noise.

Interval-based classification

The most typical interval-based algorithm would be the time series forest [22], a type of random forest. A forest is a collection of decision trees. Each decision tree is a machine learning model that takes a data point and assigns a class based on the value of its variables, as shown in Fig. 5. By creating a number of trees, each which gets to see a different part of the data, you end up with an ensemble of different tree classifiers (i.e. the forest), each possibly producing a different classification (since it has been trained on slightly different data). You then follow the usual ensemble procedure and choose the most common prediction as your final prediction.

Fig. 5. Decision tree example. A decision tree takes a data point, in this case a game, which can be characterized along a number of different dimensions (here 3: a. type of game, b. multiplayer via Wifi, c. works for 2 players). By traversing the tree, we end up at a purchase decision (as illustrated by the dashed line, a computer game that does not allow for game play via Wifi will not be bought). In a random forest ensemble the aim is to produce trees with different structure whose errors are uncorrelated.

The special thing of a time series forest is that you divide your time series into n distinct intervals and for each interval you calculate the 3 values mean, standard deviation and slope [21]. Thus you end up with 3n variables that you input into the model. One of the tricks is to find how you split your time series into intervals. The performance of time series forests are not as good as BOSS or COTE [13].

Deep learning

Deep learning models are types of neural networks consisting of several, often many, layers of neurons (See [23] for a simple introduction). They typically contain many more parameters than other machine learning models. Often they consist of a “body” of layers to which data is fed. Whereas early layers represent basic shapes (e.g. a horizontal line in image models), the later layers encode semantically meaningful representations. Thus, the output layer of the body (the “neck”?) should contain what is meaningful and salient in the data. For example, in image classification the body can capture the presence of a car wheel and the neck all the information needed to discriminate between different makes, models and versions. On top of that is a “head” which is a classifier of some sort that makes a classification based on the output of the neck. By working on the neck rather than directly on the image, it is easier for the head to classify data into e.g. cars vs airplanes. In the last decade, deep learning models have become the obvious choice for classification of both images and language. For time series it has only recently gained in popularity somewhat, but still performance on the UCR/UEA datasets is as good as HIVE-COTE [31924] despite being much faster.

Ways of training deep learning models

There are perhaps two different ways to train a deep learning model for time series classification (as is the case for many types of deep learning): unsupervised (generative) and supervised (discriminative) [23]. Supervised is what we talked about above. This is where you train the model to generate the correct classification. One problem with this approach is that the classes contain only little information (for a two-class problem only 1 bit of information), whereas the time series contains a lot more and the model has many parameters to be tuned. With generative learning one instead trains a model to output a whole time series. One approach is the auto-encoder, where one replaces the head with an inverted body. Since the neck contains a distillate of the information of the time series, by passing it through the inverted body one should be able to regenerate the original time series. Since the time series contains more information than the class, the network is trained to capture all the information in the neck that is needed to regenerate the time series well enough. Having performed basic training, one can then remove the inverted body and replace it with a head that you can then train on the classification task at hand. TimeNet is an example of this approach [25].

Model types

There are many different types of deep learning models. Since the field moves quickly in this area, I will not go into detail concerning the specific architectures. They are generally one-dimensional, and often are causal (meaning that a value only depends on historical values, not future ones). Previous models were often of a type called recurrent networks, but recently the most successful architectures have been so-called convolutional networks. Some of the architectures that perform at least as well as COTE are ResNet [3], Fully Convolutional Networks [3] (that can handle time series of different length), multi-scale convolutional neural networks [3] and InceptionTime [26].

Lack of data problem

Lack of data might be the main reason why training from scratch of deep learning models tailored specifically for time series data has not previously been able to beat the more traditional algorithms on the UCR/UEA datasets. This is suggested by the big discrepancy between time series and image dataset sizes [6] as well as the improvement of classification accuracy with deep learning model ResNet when synthetic data is added to the smallest datasets in UCR/UEA [27]. We end this section with some approaches to tackle the data problem.

ROCKET [19] is a model (not a typical neural network but with many similarities) that has taken the radical approach of randomly setting the weights of the model body without training. It chooses a large collection of different weights of various magnitudes in the hope that they will be able to capture a large variety of patterns in the data. This model has very similar performance to HIVE-COTE.

Another approach has been transfer learning. TimeNet [25] is created specifically to serve as a pretrained and publicly available time-series specific deep learning network for common use. There is also the transfer learning from image models mentioned above.

Finally, a third approach has been to increase the amount of data by producing artificial data. A good overview is given in two articles by Iwana and Uchida [628]. This technique is called data augmentation and merits a blog post of itself, but one of the approaches has been to take two time series from the data belonging to the same class and mixing them. The secret of the sauce is how to do the mixing properly. As when calculating the distance between time series, one needs to find out which pairs of data points in two series are corresponding. So one approach has been to use DTW to pair up the time points of two different series and then take an average of the paired data points. Currently data augmentation is not an established practice in time series classification whereas it is in image classification [628], but this is a field that will become much more important in the future. It will probably boost the performance of most algorithms, but especially more complex models such as deep learning models.

Last words

Above I have tried to give an outline to the enormous field of time series classification. For readers, whether you do or do not know machine learning, I hope that the above can guide you and set expectations should you start a time series classification project. One suggestion would be to find a dataset from UCR/UEA [1] that is similar to yours in order to get an idea of what could be achieved. If you know how to code and want to try out an algorithm, I suggest you download a dataset from UCR/UEA and try out a basic algorithm such as DTW+1-NN ([29] for a Python implementation) or one of the algorithms that perform well on a suitable UCR/UEA dataset. There are many algorithms out there, and many of the articles reviewed above come with source code from Papers with code [30]. Note however that the code is written by researchers and not professional developers, which means that the code might require some tweaking.

Working on a machine learning project is similar in many ways to working on a code development project. It is great to try things out on your own. Still, just like most code development work sooner or later needs developers, the same goes for machine learning projects.

References

  1. UEA & UCR Time Series Classification Repositorytimeseriesclassification.com.
  2. Schäfer, P. (2015) The BOSS is concerned with time series classification in the presence of noise. Data Min Knowl Discov 29, 1505–1530.
  3. Wang Z. et al. (2017) Time Series Classification from Scratch with Deep Neural Networks: A Strong Baseline. IJCNN 2017: 1578-1585.
  4. Cover T.M., Hart P.E. (1967) Nearest neighbor pattern classification IEEE Trans Inf Theory 13(1)21–27.
  5. Müller M. (2007) Dynamic Time Warping In Information Retrieval for Music and Motion 69-84
  6. Iwana B. K., Uchida S. (2020) Time Series Data Augmentation for Neural Networks by Time Warping with a Discriminative Teacher. arXiv:2004.08780v1.
  7. Sultana F. et al (2018) Advancements in Image Classification using Convolutional Neural Network 4th Int Conf Res Comput Intell Commun Netw (ICRCICN), Kolkata, India, 122-129
  8. Krizhevsky A (2009) Learning Multiple Layers of Features from Tiny Images.
  9. Imagenetwww.image-net.org.
  10. Wang Z., Oates T. (2015) Imaging Time-Series to Improve Classification and Imputation. In: 24 Int Joint Conf Arti Intell.
  11. Dau H.A. et al (2019) The UCR Time Series Archive IEEE/CAA J Autom Sinica, (6) 6, 1293-1305
  12. Ye L., Keogh E. (2009) Time series shapelets: a new primitive for data mining. In KDD ’09: Proc 15th ACM SIGKDD Int Conf Knowl Discov Data Min.
  13. Bagnall, A. et al. (2016) The great time series classification bake off: a review and experimental evaluation of recent algorithmic advances. Extended Version. arXiv:1602.01711.
  14. Amidon A. (2020) A Brief Survey of Time Series Classification Algorithms Towardsdatascience.com.
  15. Gupta A. et al (2020) Approaches and Applications of Early Classification of Time Series: A Review arXiv:2005.02595v2.
  16. Bagnall A. et al. (2015) Time-Series Classification with COTE: The Collective of Transformation-Based Ensembles. In IEEE Trans Knowl Data Eng., 27, (9) 2522-2535.
  17. Lines J. et al. (2016) HIVE-COTE: The Hierarchical Vote Collective of Transformation-Based Ensembles for Time Series Classification. In IEEE 16th Int Conf Data Min (ICDM), 1041-1046.
  18. Shifaz A. (2020) TS-CHIEF: A Scalable and Accurate Forest Algorithm for Time Series ClassificationData Min Knowl Discov 34, 742-75.
  19. Dempster A., Petitjean F, Webb G.I. (2020) ROCKET: exceptionally fast and accurate time series classification using random convolutional kernels. Data Min Knowl Discov 34, 1454–1495.
  20. Lin J.et al (2007) Experiencing SAX: a novel symbolic representation of time series. Data Min and Knowl Discov, 15(2):107–144
  21. Lin J., R. Khade, and Y. Li. (2012) Rotation-invariant similarity in time series using bag-of-patterns representation. J Intell Inf Sys, 39(2):287–315
  22. Deng et al. (2013) A time series forest for classification and feature extraction. Inf Sci 239, 142-153.
  23. Kaggle Intro to Deep Learning.
  24. Fawaz H.I. et al (2019) Deep learning for time series classification: a review Data Min Knowl Disc 33, 917–963
  25. Malhotra P. et al. (2017) TimeNet: Pre-trained deep recurrent neural network for time series classification In: ESANN 2017 proc, Eur Symp Artif Neural Netw Comput Intell Mach Learn Bruges (Belgium).
  26. Fawaz H.I. (2020) InceptionTime: Finding AlexNet for time series classification Data Min Knowl Discov 34 1936–62
  27. Fawaz H.I. et al (2018) Data augmentation using synthetic data for time series classification with deep residual networks arXiv:1808.02455.
  28. Iwana B. K., Uchida S. (2020) An Empirical Survey of Data Augmentation for Time Series Classification with Neural Networks. arXiv:2007.15951v1
  29. Regan M. (2018) Github code repo: K-Nearest-Neighbors-with-Dynamic-Time-Warping
  30. Papers with code – time series classification

Fredrik Edin
Konsult Machine learning & data science


I learned React in 3 weeks and so can you!

Let’s start right away with the premise:

If you have few years of experience in software development, you can learn any programming language or framework in 3 weeks.

In 6 to 12 months no one will ever suspect that this is not your main language. Not your coworkers, not your manager, not even your own mother! Any language you pick today can be added as a skill to your LinkedIn profile 3 weeks from now!

Disclaimer

While the target audience for this article is mainly front-end developers with at least few years of experience, it is not really a technical one. Despite the abundance of software terminology and abbreviations, it is in fact a motivational one. I encourage you to read through it and skip the parts that make no sense to you. Happy reading!

For 10 years I’ve been doing Android almost exclusively. OOP was my religion. I had Clean Code under my pillow and a picture of Uncle Bob on my wall. Sure, I’ve switched from Java to Kotlin in 3 days, but Kotlin is pretty much the same, just fancier, synthetic sugar without calories. And sure, I’ve written some scripts to batch rename files or automate birthday wishing to my mom, but never Java Script, not even a single line. That would be blasphemy. Utter chaos. Never even did Right Click -> Inspect Element. The web is a black box to me. No hablo inglés.

In this article we’ll go through 5 specific steps that will make you an “expert” in the language you choose to learn. Expert of course has quotation marks, but it is also not an over statement. You will create a decent solution. It will look like is done by an expert. It will feel like is done by an expert. Why you may ask? Check this photo below:

Now, you don’t need to be an architect or an interior designer to understand that this is not a proper door placement, it is a disaster waiting to happen. Similarly, you don’t need to be an artist to appreciate a beautiful piece of art. Even if you don’t know details about light, shapes, color, composition and perspective, millions of years of evolution shaped our brains to appreciate the beauty of aesthetics. It is embedded in nature. It is universal.

So, if you ever wrote a piece of code that you thought it is beautiful in your primary language, you will write somewhat equally beautiful code in the language you will learn. Not just that, you will also be able to recognize beautiful code when you are googling stuff. You will be selective in your copy pasting.

Before we dive right into the 5 steps, let us first tell:

A short backstory

Few months ago I started my own company. As an entrepreneur, the most important decisions you will make is the choice between I will do this myself for a fraction of the cost and I will pay for a service. Few examples:

  • Build your own website or use Wix
  • Deploy your own backend or use Firebase
  • Do your own paperwork or use Red Flag
  • Create your own content or use GPT-3

While #1 and #2 are a matter of preference and largely depend on your existing skill set, #4 is a human’s denial that creativity is related to consciousness and there is no way that an algorithm can easily replace us, #3 is something absolutely no one wants to do!

Your idea is going to change the world and you cannot change the world if you are stuck filing in the receipt for the new PS5 console for your kids as a business purchase. You don’t want to learn the tax system, so you hire an accountant to deal with all that.

Soon, you discover that accountant also deal with expensive, clunky legacy systems that require a great deal of manual work. Pasting stuff in excel sheets, organizing folders on cloud storage, exporting reports in obscure formats. Naturally, you want to help with all of that, you want to provide a solution that will make life a bit easier for your accountant and find your first client in the process. Business is fun!

1. Find your motivation

The backstory leads us to step number one: Find your motivation. This is imperative for your learning process. Following basic tutorials or reading articles about best practices is a good place to start, but in order to truly progress, you need to get your hands dirty working on a real life problem, providing a real life solution.

So, go ahead and find your motivation. Your pet project that needs a revamp, that app your cousin insists is going to be the next unicorn, a simple tool to connect with your grandma that features two big emojis and a parallax header of you being little and eating her meals, a beautiful app featuring a map and a gallery of misplaced electric scooters,… Anything works.

Not motivated by random examples? Just ask a person you interact with about a problem they have and try to solve it with software. Make someone redundant. Have fun.

Still nothing? There is an app for that! You can even build an app idea generator yourself.

Bonus step: Find your time

Motivation is nice and all, however you do need to find time to express yourself creatively. If your partner or child is screaming for attention while you navigate tight deadlines, maybe learning a new language will add fuel to the fire. But on the other hand, you get ahead in life by over promising and hopefully over delivering, so feel free to skip this step, who needs time!

2. Understand your motivation

Congratulations, you found your motivation and have a nice thing you can build. Now go ahead and light a scented candle, play some relaxing music, assume a meditative posture and look deep inside yourself to understand why are you doing this.

Few common reasons:

  • You are bored during the pandemics
  • You feel like you are lagging behind with your skillet and feel a need to stay relevant
  • You are helping out a family member, a friend or an associate
  • You secretly hope that this will scale one day, you will become ultra rich and beat the CEOs of tech companies in their quest to colonize the solar system
  • You accept the reality that no human will ever need to code 10-ish years from now and you want to enjoy it while it lasts

Understanding the reason behind your willingness to learn a new language or a framework will help you tremendously in making the right decisions. It will do this by revealing the constraints.
For example, to solve the problem my accountant was facing we needed to make a solution that is:

  • free or extremely cheap – we try to optimize, not introduce more cost
  • multi platform – mobile apps and a web admin tool
  • complete – you are trying to solve a problem after all
  • doable within a month or two – your time is precious
  • not terrible – functional and stable, but not too fancy

And this is how I learned React. Found a reason to build something, understood what needs to be done and chose a framework that covers most of the ground. With little difference between React Native and React JS you can cover the mobile apps and the web admin tool by learning a single language (JavaScript) and single framework (React). You can even take it up a notch and reuse some of the code, but let’s not get ahead of ourselves.

Bonus step: Doubt your understanding

Why React and not Flutter? You can pretty much build web applications with Flutter as well, plus Dart is much closer to Java your old friend than JavaScript. Oh, I just got reminded how much I miss Kotlin. Synthetic sugar overload. So slick. How is Kotlin Multiplatform coming along? What about Ktor? Too many options to choose from. Information overload. We are providing a solution, not waging tech holy wars. Skip this step, React is just fine.

3. Start from the top

Contrary to the popular belief, masterfully expressed in one of the Drake’s songs,

you really don’t need to start from the bottom. You don’t need to understand the basics of JS syntax, the difference between undefined and null or flexbox magic. You don’t even need to dread the lack type safety and consider Type Script. None of that. You only need to add the current year to your web searches. For example: “How to build a React Native App 2020” or “Free web admin templates React 2020” and start from there.

React introduced hooks in 16.8 so goodbye classes! You will never ever have to learn that part of history, unless for some reason you decide to work on legacy code. What about state management? Redux seems like a super boilerplate-y answer to your needs. Add 2020 to that search and you’ll discover a toolkit. Start from there. What about Recoil. Maybe none of that. Hooks will do the job.

What about UI that looks consistent on all platforms? Check out Material or Paper, all apps and all web pages look alike nowadays, no need to reinvent the wheel. Everything that you want to achieve is already done, you just need to put it together. Oh wait, Adobe made Spectrum. By the time this article gets published 20 more frameworks will pop out. It is obsolete already.

Anyways, start from the top and google your way down. It is impossible to understand every single concept you will encounter, every underlying mechanism or language sugar. It is also unnecessary, you have your mission, just search for the thing you are trying to figure out, add the current year to the query and read away.

Bonus step: Go strict

Establish a strict policy. Consider lint warnings as errors. Make your code pretty. Add style hooks to your commits. If you are starting from zero might as well go all in. In the next 3 weeks you will practice expressing yourself in a different manner, it only makes sense to try to express yourself flawlessly. Or it doesn’t. Beauty is in the eye of the beholder. Skip this step, it introduces too much friction.

4. Connect with your inner child

You know who is especially good at learning? Kids. They are awesome at learning. No wonder half of the spiritual advice you can get nowadays is to connect with your inner child if you want to grow and experience the world in a more profound way.

Kids don’t understand the rules of football. They have no idea what offside is. They have a ball and a goal and they kick the ball in the general direction of the goal. Over time they learn all known tricks that increase the likelihood of having the ball in the goal and over even more time they introduce new, previously unheard of tricks that get them ahead in the game.

This is the best way to approach this. Not just this, anything new you want to learn in life. You don’t need to understand everything or discover all the tricks, you just need to aim for the goal and commit to kicking the ball towards it. A little bit closer every day.

Connect with your inner child. Explore your curiosity. Ask questions, but don’t stop when you don’t get a proper answer. Be undefined. Choose what value you want to assign yourself as you go. Don’t let a static compiler check to determine who you are. Play. Have fun. Make bold claims and amend them the next day. By the time you are close to a shooting distance, you can scissor kick the ball, publish your app, update your LinkedIn profile with the new skill and write a blog post explaining how you shot for the stars and landed on the moon.

Bonus step: Scrape everything and start over

OK, don’t do that just yet. Scrape everything and start over. Your code is bad. It will never scale. Your view layer is tangled with your business logic. You don’t even downscale the images you upload to your storage and will soon reach the quota limits of your free plan. People will hate it. There are much better solutions out there. You were right from the start to doubt yourself. Again with the overthinking. Is this a part of the normal development process? Silence the internal critic and skip this step as well, you made something awesome!

5. Go public

You have to. Going public is the best way yo make a commitment. You don’t have to publish the app on the App Store or Google Play and go through cumbersome submission processes, but go public somehow. Send an email to your friends asking for feedback, write a semi-humorous blog post to justify your efforts, push your code to GitHub, answer a question on Stack Overflow with your new gained expertise, make a PR on the library you used, motivate your surrounding to learn something new as well,…

Get creative, contribute somehow to the big data, make it easier for the algorithms to understand how humans learn, they may suggest improvements in the future based on your experience.

You may feel a strong urge not to go public, especially if you did not manage to produce something tangible in these 3 weeks. But, before you go ahead and associate work with measurable output, ask yourself these two questions:

  • What did you learn?
  • Did you have fun?

If you learned something you did not know before and you had fun doing it, than you’ve successfully completed this tutorial and you’ve proven yourself that what sounds improbable is in fact quite possible. You are awesome, give yourself some love

A short summary

That’s all. Well, probably there is lot more to it, but this is the level of insight you can expect from a random blog post on the Internet about learning something that differs only slightly from what you already know. Not a lot of new information, nothing too profound or insightful.

Just a gentle reminder that you are not defined by your past choices. You are not limited by your existing knowledge. You are not stuck with a technology. You are free. You can learn pretty much anything at any time and have an amazing time doing it.
So, go ahead and do it!


Medo is an entrepreneur with extensive experience in software development and team leadership. He engages in projects ranging from hacked together proof of concepts to distributed solutions serving millions of users.

Medo loves sharing, he believes it is the optimal way of learning. He often speaks at events and social media on topics varying from technical solutions to inspirational talks. His enthusiasm for exchanging knowledge goes from vegan food recipes and body hacks, to quantum physics and the origin of the Universe.

At the moment, he is pursuing a dream to create a better workplace for everyone, helping individuals and organizations embrace a more mindful and heart based approach to leadership. He would love if you connect with him on LinkedIn.

Ivan Dimoski
Android expert & Mindful leader


Elixir och en framtid med större möjligheter

Elixir är ett programmeringsspråk som vunnit mycket mark på senare tid och i min erfarenhet vinner ovanligt mycket mark bland mer erfarna, seniora, utvecklare. Jag vill belysa lite av de systematiska fördelar jag ser Elixir ge företag och utvecklare. Det blir ingen kod, bara översikt och förklaring i det här läget. För att kunna bekanta oss med Elixir och dess fördelar behövs en liten historielektion som ger oss Elixirs tekniska arv och sammanhang.

​Elixir bygger på arbetet med ett annat språk som heter Erlang. Det är ett språk som utvecklades på Ericsson på 80-talet och därefter släpptes som open source. Det skapades för att bygga system med fokus på driftssäkerhet, tillförlitlighet. Mer specifikt byggdes det för att ge mjuka realtidsgarantier för distribuerade system inom telekomsystemen som Ericsson skapade. Erlang har ovanliga egenskaper som man inte ser i särskilt många andra språk. Särskilt inte samlade:

  • Hot code reloading: Man kan uppdatera koden i en körande applikation och bibehålla existerande tillstånd, utan omstart.
  • Stark introspektion under pågående körning.
  • Mjuk realtid, tillförlitliga latency, preemptive scheduling: Under körning kan VM:en avbryta pågående arbete för att säkerställa att alla pågående processer får chans att köras på CPU:n. Detta hindrar flera patologiska feltillstånd, t.ex. oändliga loopar, från att krascha eller låsa applikationen.
  • Distribution & klustring: Erlang har en inbyggd lösning för distribuerad körning och klustring som bygger på message passing. Detta grundar sig på att Erlang är byggt enligt Actor-modellen.
  • Concurrency & Parallellism (använda flera kärnor) Väldigt få språk kan effektivt använda CPU-kärnor hejvilt men i Erlangs fall så hanteras det med samma modell som utvecklades för distribution. Lättviktiga “processer” istället för trådar. Immutability och message passing för tråd-säkerhet. Så det är billigt och tämligen ofarligt att nyttja flera kärnor i en applikation, utan att använda en worker-pool i någon slags applikations-server eller köra fler instanser av applikationen.

​Erlang har aldrig varit särskilt populärt men har länge respekterats i sin niche. För vissa företag har det varit ett hemligt vapen. Det är skapt för att kunna bygga extremt tillförlitliga system med goda prestandagarantier och det har såvitt jag vet alltid levererat väl på det området. Det har förstås också utvecklats löpande sedan 80-talet. Och det har körts i seriösa driftsammanhang sedan dess.

​Elixir ger Erlang en fräsch och enklare syntax. Det lägger vidare till en hel del smidigare funktionalitet för modern utveckling och har blivit en renässans för Erlang och BEAM:en, den VM som Erlang och Elixir körs på.

​Så varför bör vi bry oss om Elixir?

Grundplåten

​Antalet språk som har ett gott rykte som exceptionellt pålitliga är få. De är också ofta trögarbetade. Elixir är ett tight och uttrycksfullt högnivåspråk som har mycket av sitt syntaktiska arv från Ruby. Elixir är dock funktionell programmering, inte objekt-orienterat. Precis som Erlang.

Webbramverket Phoenix, Ruby on Rails andliga efterträdare

​Elixir skapades av José Valim som tidigare varit en aktiv Ruby-utvecklare. Rails är ett extremt populärt webbramverk. För Elixir är Phoenix det motsvarande ramverket och en stor mängd utvecklare från Ruby och Rails har rört sig över till Elixir och Phoenix.

​Det går att skriva mycket om Phoenix, det är ett kraftfullt och lättarbetat webbramverk. Utifrån det så har även en del mycket imponerande tekniska lösningar byggts. En intressant utveckling är t.ex. tillägget LiveView som låter utvecklare skapa mycket dynamiska lösningar utan Javascript. Den här videon från Chris McCord kan vara en bra dragning.

Ekosystemet

​Språket har ett starkt växande ekosystem och de mjukvarubibliotek som behövs finns i allmänhet i mogna implementationer. Det har även fördelen av att kunna luta sig på och använda bibliotek som skapats för Erlang när Elixir-varianter helt saknas. T.ex. bygger Phoenix på en HTTP-server som heter Cowboy som utvecklades för Erlang.

IoT & Connected Devices

​Inom Elixir har ett projekt som heter Nerves skapats som har ett exceptionellt ambitiöst och synnerligen kapabelt sätt att närma sig arbete med hårdvara för IoT och liknande embedded-projekt. Om det kan köra Linux så kan det tjäna på Nerves. Det är en verktygskedja med exceptionellt snabb iterationshastighet för hårdvarunära utveckling. Det nyttjar en kombination av gedigen erfarenhet av hårdvaruutveckling och Erlangs egenskaper för att skapa pålitliga hårdvaruenheter. Det har även skapats ett verktyg vid namn NervesHub som löser säkra uppdateringar av produkter som är ute hos kunder i fältet vilket kan vara väldigt komplext. Det går att läsa mer här.

​Ett annat ramverk som gör nytta inom IoT och Connected Devices är Scenic. Detta är ett UI-ramverk som använder OpenGL för att rendera lättviktiga UI med ett starkt säkerhetstänk och god flexibilitet. Scenic är enormt lättviktigt jämfört med den populära lösningen att bädda med en hel webbläsare och webbapplikation. Det är få system som har egna lågnivå-ramverk för att rendera användargränssnitt och Scenic är en imponerande lösning för det. Här presenteras det på ett talk på Code BEAM i Stockholm.

Varför Elixir?

​Så varför bör vi överväga Elixir?

​Med en stark modell för concurrency och parallellism drar det bättre nytta av serversresurser än genomsnittliga applikationer skapta i t.ex. Python, Node.js, Ruby, PHP och många fler. Detta sparar pengar, ofta rejäla pengar, på hårdvara eller notan hos molnleverantören.

​Det är en stark grund att stå på. Elixir och Erlang erbjuder verktyg som i princip inga andra språk eller ramverk rimligt kan erbjuda när det gäller slagtålighet i drift, distribution och skalbarhet,

​Funktionell programmering är enligt många avsevärt enklare att resonera kring. Min erfarenhet som konsult som klivit in i andra organisationers Elixir-projekt är att jag är feature-produktiv dag ett eller två och blir bekant med systemet snabbare. Vare sig jag återbesöker kod jag skrev för 6 månader sedan, dyker ner i koden i ett open source-bibliotek så är den lättare att plocka upp då jag inte behöver hålla en klass-hierarki och objekt-state i huvudet när jag läser och skriver kod. Input, output och minimalt med bieffekter.

​Den största utmaningen är väl just att funktionell programmering är ett annat paradigm än objekt-orienterat och att objekt-orienterat fortfarande är dominant. Men sen har ju många ekosystem, särskilt Node.js/Javascript plockat upp funktionella koncept så det är mer och mer bekant för utvecklare.

​Man kan bygga webbsystem och backends i nästan vilket språk som helst och få det att fungera fint. Det är min bestämda uppfattning att Erlangs VM, BEAM, erbjuder egenskaper som man inte hittar i något annat moget ekosystem eller språk. Och att de egenskaperna är enormt användbara för majoriteten av backendsystem som byggs idag. Elixir är nyckeln till att göra Erlang lika enkelt som andra moderna språk. Och därav ser jag att Elixir, till sin största fördel, höjer taket för hur bra ett system kan byggas.

​Mail: lars@underjord.io Twitter: @lawik Webisite: underjord.io

Lars Wikman
Rådgivare & problemlösare


Introduktion till Flutter för utvecklare

Varje gång ett nytt ramverk, programmeringsspråk eller verktyg presenteras så blir vi lovade att detta kommer lösa alla problem som de tidigare alternativen misslyckades att hantera. Flutter, Googles cross-platform lösning för appar, är egentligen inte annorlunda ur den aspekten. Det som talar för Flutter handlar mer om att Google undvikit de misstag som andra cross-platform lösningar gjorde, och på så vis har de skapat något som jag tror kommer ta över en stor del av app-utvecklingen framöver. 

Oavsett om jobbar med backend, frontend, eller apputveckling, så är Flutter en teknologi som du borde känna till och förstå. Det som gör Flutter populärt är sådant jag är övertygad om kommer ha en inverkan på andra tekniker. Genom att förstå Flutter så blir du alltså bättre förberedd på hur både appar och webb kommer att byggas framöver.

Deklarativt UI

På senare år så har konceptet declarative programming dykt upp som ett populärt alternativ till det mer traditionella imperative programming. Jag kommer inte göra någon djupdykning i skillnaderna mellan dessa, utan istället fokusera på hur deklarativ programmering fungerar i Flutter. Min förhoppning är att min genomgång och exempel här ska få en bild om varför det är en stor fördel framför andra modeller, och varför Flutter också är ett bättre alternativ än andra deklarativa ramverk (sett till vad vi har för alternativ idag).

Flutter deklarativa modell för att bygga användargränssnitt bygger på Dart, det programmeringsspråk som Flutter använder. Genom en smart design av hur koncept som optional parameters och default values fungerar så har man kunnat skapa en modell för att beskriva användargränssnitt på ett konkret och lättläst vis. Andra ramverk, som React, har också en deklarativ modell, men där har man fått kompromissa med underliggande tekniker och befintliga standarder på ett sätt som skapar en mindre optimal modell för att bygga gränssnitt. Följande kod är ett exempel på hur man bygger en enkel widget i Flutter:

Ovanstående visar på hur Flutters deklarativa modell uppnås genom språket Dart. Ett antal olika widgets kombineras i ett träd för att komponera vår nya widget. Detta görs genom att i funktionen build() konstruera vårt träd. Column representerar helt enkelt en kolumn med flera andra widgets under sig. Vi skapar denna genom att anropa konstruktorn och skicka med två parameterar, mainAxisAlignemnt och children. Den första, mainAxisAlignment, är en valfri parameter och om den inte anges så får den värdet start, vilket medför att dess barn positioneras från toppen och nedåt. Den här modellen att ha valfria parametrar med standard-värden gör att man oftast bara behöver ange ett fåtal när man använder en befintlig widget, vilket i sin tur förenklar koden man behöver skriva.

I Flutter är widgets inte bara komponenter som är synliga, utan också komponenter som lägger till utfyllnad (padding) eller som asynkront laddar data och bygger om din widget när den är klar. Följande exempel visar hur widgeten FutureBuilder kan användas för att ladda data asynkront och visa en ikon (med en smiley) innan den är klar, för att sen bytas ut mot en Text med användarnamnet.

Observera att en widget alltså kan välja att rita om delar av sig själv och behöver inte ens returnera ett träd med samma typ av widgets när det sker. Det här gör det alltså möjligt att beskriva vårt användargränssnitt baserat på olika villkor, och fortfarande använda en deklarativ modell.

Developer Experience

Ett begrepp som har börjat användas mer på senare år är Developer Experience. Detta används för att beskriva hur användningen av ett utvecklingsverktyg, programmeringsspråk eller ramverk upplevs av utvecklare. En bra Developer Experience är alltså något som utvecklare uppskattar och upplever vara mer effektivt att arbeta med. Huruvida en teknik är bättre eller ej handlar ju ofta om subjektiva uppfattningar, vilket är anledningen till att en god Developer Experience är en viktig komponent när man ska välja teknik. Är verktygen undermåliga, programmeringsspråket föråldrat, eller ramverket komplicerat att använda så är risken stor att det kommer medföra att utvecklarna blir mindre effektiva.

Flutter har ett utmärkt stöd i den mest populära texteditorn på marknaden idag: Visual Studio Code. Det finns även tillägg till IntelliJ IDEA från JetBrains om man föredrar det verktyget. Bra stöd är viktigt, men för att bli riktigt framgångsrikt så måste verktygen också vara smarta. Flutter-stödet i Visual Studio Code (och IntelliJ IDEA) har flera viktiga funktioner för att snabbt kunna redigera koden för ett UI. I exemplet nedan visas hur man snabbt och enkelt kan modifiera sitt widget-träd med hjälp av enkla snabbkommandon.

Denna funktion blir extra viktig när man använder en annan funktion i Flutter: Hot Reload. Detta låter en utvecklare uppdatera en app som körs på en telefon eller emulator, utan att behöva starta om den helt. Det går alltså att i realtid se hur ditt UI påverkas bara genom att du sparar koden i texteditorn efter en ändring.

Smarta funktioner ger en bättre upplevelse för utvecklarna. Detta gör dem i sin tur mer effektiva när de arbetar med den tekniken, vilket reducerar tiden det krävs för att bygga klart appen. Developer Experience är en viktig komponent, och Flutter är idag bäst i klassen.

Fokus på prestanda

Ett vanligt problem med cross-platform lösningar är att de måste kompromissa på områden som gör att slutanvändaren kan bli lidande. Ett vanligt problem är att prestandan i användargränssnitten blir avsevärt mycket långsammare än om de skrivits med den native teknik som respektive operativsystem erbjuder.

Flutter har från början lagt ribban för prestanda riktigt högt. Motorn i Flutter garanterar minst 60 FPS (120 FPS om hårdvaran tillåter det), och det är svårt för en utvecklare att av misstag strypa den hastigheten. Detta gör att användarupplevelsen i en app byggt med Flutter känns rapp och responsiv utan att användaren behöver lägga särskilt stort fokus på det området.

Prestandan i Flutter handlar inte bara om hur många FPS den kan rita ut i, men också om att den kan bygga om ett stort och relativt komplext träd av widgets mellan två frames. Det är alltså inget större problem att ditt UI helt ändrar struktur och bygger om alla widgets, då motorn i Flutter är optimerad för att göra just detta. Det här gör att koden för en Flutter app blir enklare och lättare att återanvända. Man behöver inte längre bygga avancerade komponenter som klarar många olika utseenden. Istället bygger man widgets som är specialiserade och växlar snabbt mellan dem vid behov.

Fokus på användargränssnitt

Även om appar har ett stort fokus på användargränssnitt så finns det många situationer när man vill att en app ska köras i bakgrunden. Vanliga exempel är musikspelare eller träningsappar, som vi vill ska fortsätta att köra även om de inte ligger i förgrunden eller när skärmen är släckt. Detta är också exempel på de typer av appar där Flutter passar mindre bra.

Eftersom Flutter har fokuserat helt på att skapa ett ramverk för att bygga användargränssnitt, så har möjligheten att exekvera kod i bakgrunden fått tagit ett steg bakåt. Även om det finns ett visst stöd för detta så är det betydligt mer komplicerat och man kommer ofta få gå över till att skriva kod i Swift respektive Kotlin för att lösa vissa mer plattformsspecifika delar.

När man ska bestämma sig för att använda Flutter eller ej så är detta en bra utgångspunkt. Ska appen köras i bakgrunden, eller direkt använda mer avancerad hårdvara så är det inte rekommenderat att använda Flutter. Appar som musikspelare, rena kamera-appar, träningsappar, appar som använder Bluetooth, o.s.v. är alltså mindre lämpliga kandidater.

Om fokus däremot ligger på själva användargränssnittet och aktiv användning så kan Flutter mycket väl vara ett bra alternativ.

Sammanfattning

Flutter ger dig en avancerad, modern, och högpresterande lösning för att bygga appar. I dagsläget så har fokus legat på iOS och Android, men ramverket har också beta-stöd för webb och desktop (macOS, Windows och Linux). Google satsar stenhårt på att utveckla ramverket och andelen utvecklare som använder Flutter växer dagligen. I dagsläget är Flutter mer populärt än React Native på både GitHub och StackOverflow.

Sammantaget så ser framtiden för Flutter väldigt ljus ut. Tack vare att det har ett stort företag bakom sig, som dessutom driver utvecklingen av den vanligaste mobilplattformen idag (Android), så lämpar sig Flutter mycket väl för många olika typer av applikationer.

Du kan läsa mer om Flutter och hur man bygger appar på https://flutter.dev

För att läsa mer av Erik Hellman kan du klicka in på hans egen blog https://hellsoft.se

Erik Hellman
Mobilutveckling, IoT, Backend/Cloud & Webb