Projects

The overview picture of operation of LoRaWAN network ([source](https://www.thethingsnetwork.org)) In this post, I will tell you about an exciting IoT device me and my team prototyped at MakeZurich week-long makerdays. Keep reading if you are a maker, IoT enthusiast or a person interested in using IoT principles in a real world scenario. MakeZurich MakeZurich is the “Civic Tech and LoRaWAN Hackdays for a better city”. It is organized jointly with the city administration to explore new ways of solving challenges the city is facing with the help of open networks and civic tech. I missed the first run, that took place last year, because I found out about the event too late. This year, I had it in my calendar and was ready to take part. The event had a smell of a hackathon: difficult and open-ended challenges to be tackled in innovative fashion in too little time in small teams, mostly with people you didn’t know before the kickoff. But it was more than that. Rather than hackathon,it were Makerdays or Hackdays. It lasted for full 8 days, and on each day a makerspace was open so that I could just drop by, ask people about technical issues I was having and hack on the prototype and code for our project. In retrospect, I view the event as great learning experience that gave me quick insight into the field of IoT and LoRaWAN. (Big thanks to the organizers for all their work, it showed!) ...

In this post, I will demonstrate and explain Reinforcement Learning code I developed. You will learn how one can train an AI agent to master Atari games and understand the technology behind DeepMind’s AlphaGo, the first computer program to defeat professional human Go player. This will be fun and, surprisingly, simple, so let’s dive in. OpenAI Gym OpenAI’s gym is a toolkit for developing and comparing reinforcement learning algorithms. Big standardized datasets have proven pivotal in development of deep learning algorithms. Similarly, collection of test problems, environments, as provided by gym aids in evaluation of reinforcement learning algorithms when an agent learns to take actions in response to observations from environment to “solve it”. All that comes below is enabled by the work of guys at OpenAI. Hat tip to them. ...

I am sure that most of you have heard about many of the deep learning libraries out there including TensorFlow, Theano, Keras and PyTorch. They facilitate building of layered and potentially complex neural network models for areas as diverse as automatic image captioning, speech recognition, and drug design. Undoubtedly and for better or worse, the deep learning field is riding on a wave of hype. The good thing about it is definitely that the deep learning libraries keep developing at astonishing pace. In this article, I will make use of these developments to implement two stochastic models from systems biology and solve them efficiently using the computational graph paradigm of Theano and TensorFlow. ...

Source Introduction Genetically engineered cells and organisms are being used to produce array of commonplace commercial products including [drugs](http://www.madehow.com/Volume-7/Insulin.html) and [materials](https://www.bio.org/articles/current-uses-synthetic-biology). Biological engineering is being employed to enhance nutrition content of various foods and yield of crops. Recently, [lab-grown meat](https://en.wikipedia.org/wiki/Cultured_meat) represents an interesting alternative to traditional production and [synthetic biology](https://www.nature.com/news/2010/100120/full/463288a.html) is being [explored even by artists](https://www.ginkgobioworks.com/2018/04/11/creative-in-residence/). These successes are, however, just baby steps in face of the richness and complexity of "products" nature is so adept in engineering. Like what you ask? Like YOU, for example. In the meantime, development is ongoing at rapid pace. We are getting finer understanding of molecular-scale processes by which the life is implemented and by extension, we are becoming more adept in exercising control over them. Biology is becoming easier to analyze and design using established engineering approaches. At different system levels, from molecular to cellular to populational, mathematical modeling and abstraction in design are becoming possible. (If this abstraction thing is too abstract for you, think about this as if in your Arduino project (like this one) you had to build and debug the microcontroller first. Before you could do that, you would have to do the same for all the resistors, transistors and other elements of the board. Instead, and luckily, you just need to have the board shipped and plug it into your laptotp to boot it up. You don’t have to think of all the possibly very intricate details that make it work and can focus on building some application on top of it. This is the power of abstraction in design when put to work.). ...

*This is a second part of an introductory post on machine learning. It is based on a practical project I worked on at ETH. Here I give you an explanation of two crucial steps in machine learning pipeline – Model selection and Validation. I then wrap up with some useful takeaways from the project, that you can eventually use in your own work and concise summary. If you missed the first part of the post, or just feel you could use a refresher, you can find it here.* Model Selection Although we have discussed quite few steps so far, they were only preparatory (albeit necessary and crucial) and it is only now, in model selection phase, that we actually get our hands on machine learning algorithms. We can further divide this phase into two steps – training and validation. Let’s look at training first. In the scope of our project we needed to deal with both regression and classification, here I will describe the latter one as it is perhaps more intuitive and again, I will illustrate it on an example from our project. Recalling the task at hand – to disambiguate between mentally healthy and sick patients from an MRI scan, we can imagine a simplified scenario where each brain scan is described by single two-dimensional vector of features. Further, assume that they form two point-clouds that are perfectly linearly separable as show in the figure: Fig. 5: Idealized example of classification task Fig. 6: More realistic example of a classification task, notice that data don't form any recognizable sub-groups We may than write down an equation of line that divides the two-dimensional space into two half-planes. Whenever we make a new observation (take an MRI of a new patient), his brain, described by two feature vector, will fall to one of those two regions and we will label it accordingly as ‘healthy’ or ‘sick’. ...

*This is an introductory post for those that don’t have yet any significant experience in machine learning. It is divided into two parts, in the first one, the one you are reading right now, I introduce you to a practical project I’ve worked on at ETH. Next, I use this illustrative example to explain two important elements of the machine learning pipeline, namely preprocessing and feature extraction. In the second part of the post that I will share with you later, I will catch up on this and give you an explanation of the remaining steps of the pipeline (model selection and validation). I will then wrap up with some useful takeaways from the project, that you can eventually use in your own work and concise summary.* *If you happen to be more seasoned in the field of machine learning, don’t worry, I will address follow-up posts to you that go more into depth with things I just touched upon here. Till then, stay tuned.* You’ve heard about it many times already. It is supposedly influencing how we cast our votes, putting people off their jobs and even making sure that the pineapples you buy in your supermarket are always fresh. Machine learning is a hot topic and you are increasingly likely to come across it not only in your work, but in your day to day life in general. ...

How did you start your day this morning? First, I hope that well. Second, I guess that you did at least one of these: made yourself a coffee, brushed your teeth or took a shower. And to assume this, I don’t have to be particularly good in guessing. This is how millions of people start their day. But these activities have more in common. All of them would be either significantly different or even impossible without water running out of our taps. Luckily, clean drinking water is an infinite resource. Or isn’t it? ...