This week in AI #4: What is my purpose?
1- And the winners are…
When it comes to Computer Vision, the art of extracting information from pixels, there are two conferences you should look for: CVPR that usually takes place in the summer and ICCV later in the year. Apart from being good excuses for a business-trip to Hawaii or Venice it’s also where all the magic happens: thousands of researchers present their latest papers which are then peer-reviewed. And you guessed it, the CVPR 2017 awards are out! The two winners are Densely Connected Convolutional Networks by Facebook and Improving the Realism of Synthetic Images of which we spoke last week, if you didn’t already you should definitely check them out!
2- Neural Network debugging 101
Truth is, behind every breakthrough there’s someone who’s spent hours trying to figure out why what should be working actually… wasn’t. Although very rewarding, especially as you tend not to make the same mistake twice, it can easily become exhausting. Besides, sometimes you don’t even know where to start. Well if you’re in the business of designing and deploying neural networks you’ll certainly be interested in this handy checklist that covers 37 debugging tips: it might not be the the exact bug you’re tracking but it’s definitely a good place to start!
3- 3D content generation
Images, sound, article summaries or even entire articles, are just a few examples of content that is being automatically generated nowadays. Even though it’s not always on par with a human-generated content it’s unquestionably getting there! And more importantly the list is growing. Take for instance the latest post by Sam Snider-Held: using a clever mix of deep-learning, blender and light-fields he was able to automatically generated 3D terrain Ghibli-style! Check out the full post for all the nitty-gritty details.
4- Racist AI
Remember Tay? The microsoft twitter bot that went from harmless experiment to racist and misogynist and was shot down in a day? Here the reason was pretty easily identifiable: it learned from its user-base and simply reproduced their behavior. But most of the times it’s not that easy to find the underlying biases in your data. For instance if a bank was to have higher standards for loans based on a certain ethnicity then an AI based solely on this data would learn to do the same. Rob Speer over at ConceptNet goes over how even the most harmless pipelines can still produce racist AI. Clearly something to keep in mind and to prepare for!
5- What is my purpose? You pass butter.
From time to time it’s good to take a step-back and wonder what it is we’re really trying to achieve. AI for AI’s sake has certainly been the driving-force behind some of the biggest achievements of the last decade but where it really makes a difference in our society is when it leaves the labs to be incorporated in real-world applications. Well, count yourself lucky! The Harvard Business Review just released a special edition on AI and what it can — and can’t — be used for! If you have 15 minutes to spare you should definitely give it a go!