Google announced that it can do a zero-shot translation from Korean to Japanese and from Japanese to Korean. They do this without ever training on these language pairs, and without first translating to another language. Most people I spoke to did not get why this was so interesting. In this post I will explain what this means, and what this…
Explaining Google’s Neural Machine Translation paper
Google recently published the paper “Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation”: https://arxiv.org/abs/1609.08144. As I got a lot of questions about this paper I wanted to explain what they do to get neural machine translation closer to human translation than they could with phrase based translation. Interesting things to read after watching the video…
Building a “Smart IoT bed”: part 2 of 5
Recently I finished a cool electronics project: internet connected lights under my bed. In five steps I will guide you from project to project, each time adding functionality to what we made. In the second part we will put a small computer in our circuit, so we are able to display any color we want to have. After week 1 you were…
Ineffective sorts: tensorflowsort
Recently sequential neural networks have been gaining popularity. To explore what they were capable of I decided to create the ineffective sorting algorithm (https://xkcd.com/1185/) “Tensorflowsort”. In this post I will explain what we can do with sequential neural networks, and how we can use them to sort an array. The basis of my sorting algorithm is the Recurrent Neural Network. The…
Building a “Smart IoT bed” Part 1 of 5
Recently I finished a cool electronics project: internet connected lights under my bed. In five steps I will guide you from project to project, each time adding functionality to what we made. In the first part I will show you how to add a cool feature to your bed by connecting six electronics parts. Whenever I wake up at night I have to walk to…
Calculating how to split your bills
Dutch people are known for splitting their bills. When frequently eating with the same group this could normally lead to an awkward situation, where everybody has to pay everybody a little bit. Most students use the site “https://wiebetaaltwat.nl/”. You create a group with the people you frequently eat with, and every time you do something with a set of these…
Connection requests from strangers on LinkedIn: accepting them the right way
Many people talked with me about this dilemma: do you accept connection requests from strangers? Accepting the invite could be a foot in the door in the future, even if you don’t know this person. However: I believe that by having a connection with someone, you would vouch for this person during a solicitation. A few month ago I solved…
BNAIC 2014: using human input to improve reinforcement learning
Using human input can be valuable for reinforcement learning. Imagine the difference between having to learn something yourself, and having a mentor hold your hand while showing you how a task is performed. In the real world humans are able to generalize their knowledge: if I die by walking into an enemy from the left, I probably also die if…
Brain visualisation
During a project in which I created partitions in the brain there was no simple way to visualise my data. When asking a teacher the best method apparently was: try to write a Matlab script. To help me and others I created a new visualisation tool to display the brain right inside your browser. To show how well our visualisation…