Slashdot Mirror
Google's AI Built an AI that Outperforms Any Made By Humans (sciencealert.com)
schwit1 quotes ScienceAlert: In May 2017, researchers at Google Brain announced the creation of AutoML, an artificial intelligence (AI) that's capable of generating its own AIs. More recently, they decided to present AutoML with its biggest challenge to date, and the AI that can build AI created a 'child' that outperformed all of its human-made counterparts... For this particular child AI, which the researchers called NASNet, the task was recognising objects -- people, cars, traffic lights, handbags, backpacks, etc. -- in a video in real-time. AutoML would evaluate NASNet's performance and use that information to improve its child AI, repeating the process thousands of times.
When tested on the ImageNet image classification and COCO object detection data sets NASNet was 82.7 percent accurate at predicting images on ImageNet's validation set. This is 1.2 percent better than any previously published results, and the system is also 4 percent more efficient, with a 43.1 percent mean Average Precision (mAP).
When tested on the ImageNet image classification and COCO object detection data sets NASNet was 82.7 percent accurate at predicting images on ImageNet's validation set. This is 1.2 percent better than any previously published results, and the system is also 4 percent more efficient, with a 43.1 percent mean Average Precision (mAP).
It can identify if something is a kitten or not with 83.4% accuracy. Sounds impressive until you realize a 3 year old can do this with 99.9% accuracy.
If the 'parent' AI kept telling the 'child' AI when it was right or wrong ...
It doesn't work that way. Each NN learns on its own, using a combination of both labeled and unlabeled data. The parent NN sets "hyper-parameters", such as the number of layers, the size of each layer, the activation function, the convolution size, dropout rate, the learning rate damping factor, the batch size, etc. Then it turns the children NNs loose on the image dataset. It then sees which hyper-parameters lead to better/faster performance, and then applies ML techniques to learn better hyper-parameters.
None of this is new. What is new, is that Google is now applying this recursively, and using AutoML to design a better AutoML. This is another step toward the singularity.
Yes, this is basically just a hyperparameter optimization system that uses gradient descent instead of a random or grid search.
What would be much more interesting to see is if you could train a system to design deep learning networks that could choose good hyperparameters for a new task, in one go.
It can identify if something is a kitten or not with 83.4% accuracy.
No. It can look at an image and correctly classify it into THOUSANDS of categories, only one of which is "kitten". It was 82.7% accurate at this. If it was trained to only distinguish "kitten" from "not-kitten", it would, of course, be far more accurate.
a 3 year old can do this with 99.9% accuracy.
A 3 year old requires 3 years of training. This system can learn in hours.
Try this: https://www.youtube.com/watch?...
AI outperforms humans. There are some tricky cases after timestamp 2:42. You may want to try them for yourself.