You’ll all remember in July Mckinsey predicted that AI could automate 45% of the activities people are paid to perform and that about 60% of all occupations could see 30% or more of their constituent activities automated.
Now Forrester is weighing in claiming AI will replace 6% of jobs in the next 5 years.
I’ve really enjoyed reading people’s opinions on Reddit and would love to get your thoughts. Send us a blog you’ve written, found or rated and we’ll share with other subscribers.
Henry Lin from Harvard University and Max Tagmark from MIT say they found the secret of NN. The problem is that more mathematical functions exist than possible networks to approximate them, but the answer to this problem is that the universe is controlled by a tiny subset of all possible functions. Therefore, deep neural networks don’t have to approximate any possible mathematical function, only a tiny subset of them.
Fifteen years after the terror attack on the twin towers, Physics Professor Neil Johnson gave a talk on how AI can be used to predict terrorist attacks. For a study, him and his research team monitored pro-ISIS groups on VKontakte, the largest online social networking service in Europe. Their research revealed patterns within the formation of terrorist groups, which could be utilized to predict real world attacks.
Blurring pictures used to be an effective way to protect people’s identity, until now: Researchers at the University of Texas and Cornell University were able to identify 71% of blurred faces and numbers, compared with a 0.2% human success rate. This number increased to 83% if the computer was allowed to guess five times.
Researchers at MIT analysed large amounts of unlabeled videos in order to train a model to predict what is going to happen next. The model can generate tiny videos up to a second at full frame rate while predicting plausible futures of static images. However, the videos look dream like rather than real.
Read through part two of TensorFlow in a Nutshell’, where you will learn more about hybrid learning methods, similar to the ones used by Google in the Play store for app suggestions and YouTube for video recommendations.
New Neural Network architecture is popping up every now and then and it’s hard to keep track of them all. To solve this, Fjodor Van Veen created a cheat sheet that contains most of the new architecture.
A latest KDnuggets poll asked its users, which algorithms they used in the past 12 months for a data science related application. The average respondent used 8.1 algorithms, the top 5 being: Regression, Clustering, Decision Trees, Visualisation, and K-nearest neighbours.
Numerous research groups have been making use of GTA to train algorithms. The reason for this is that ML requires huge quantities of curated data, which can be challenging to collect. Games like GTA on the other hand already provide these datasets that can be put to use.