Predictive Artificial Intelligence
Predictive Artificial Intelligence is a sub-branch of Artificial intelligence that deals with predicting possible human responses to a given situation. Examples of Predictive AI are programs that suggest replies to email messages, programs embedded in self-driven cars that decide if a pedestrian is going to change directions, or programs that analyze posted photos to predict possible suicide attempts. These programs are based on algorithms from Predictive Analytics, Image analysis, motion analysis, and Natural Language Processing.
Examples[edit]
Responding to Emails[edit]
Google has added Smart Reply to their email and messaging apps. "Smart Reply scans the text of an incoming message, and suggests three basic responses the user can tweak and send."[1] A Long short-term memory network was trained using 300 million emails messages and their responses, keeping only the most common hundreds of thousands of n-grams (sequences of n words) to train and test on. Engineers could "only inspect aggregated statistics on anonymized sentences that occurred across many users and do not identify any user."[2] The neural network would predict the best cluster of responses and then responses that were too general (e.g., "yes!") were rejected and the results diversified to present three relatively different short answers to an email containing a question. The email user can choose to use on the predicted responses or not.
Pedestrian Behavior[edit]
Self driven cars incorporate three modules to be able to drive in the real world: a perception module composed of cameras and Lidar that detect the and identify objects around them; a prediction module that foresees where the identified objects will move to; and a decision module that controls the car's acceleration and braking.[3] Once an object is identified as a pedestrian, the onboard AI system must predict whether the pedestrian is stopped, walking, or ready to step into the road. The state-of-the-art in mid 2018 is to use a Convolutional neural network to identify a pedestrian in a scene, identify the key articulations in the pedestrian, and to feed the changes in these articulations over 10 or so frames into a trained support vector machine to determine "situations such as crossing vs. stopping, bending, and starting."[4] These techniques can predict the pedestrians action 750 ms ahead of time, giving cars extra time to stop or alert the driver. More complicated predictions of groups of pedestrians can assign Long short-term memory networks to each identified pedestrian, and then pool the results of each pedestrian to predict the possible trajectory of a given person.[5]
References[edit]
- ↑ Vincent, James. "Smart Reply is coming to Gmail for Android and iOS". the Verge. Retrieved 17 September 2018.
- ↑ Kannan, Anjuli; et al. (2016). "Smart reply: Automated response suggestion for email" (PDF). Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining: 955–964. Retrieved 17 September 2018.
- ↑ Standage, Tom (9 May 2018). "Why Uber's self-driving car killed a pedestrian". The Economist. Retrieved 17 September 2018.
- ↑ Fang, Zhijie; et al. (17 October 2017). "On-Board Detection of Pedestrian Intentions". Sensors (Basel). 17 (10): 2193. doi:10.3390/s17102193. Retrieved 17 September 2018.
- ↑ Alahi, Alexandre; et al. (2016). "Social LSTM: Human Trajectory Prediction in Crowded Spaces" (PDF). Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition: 2961. Retrieved 17 September 2018.
 | This science article is a stub. You can help EverybodyWiki by expanding it. |
This article "Predictive Artificial Intelligence" is from Wikipedia. The list of its authors can be seen in its historical and/or the page Edithistory:Predictive Artificial Intelligence. Articles copied from Draft Namespace on Wikipedia could be seen on the Draft Namespace of Wikipedia and not main one.
