 | After analyzing 15,000 samples in the dataset, we noticed that increasing the number of images doesn't significantly improve the scoreboard recognition quality for our neural network. However, what's more interesting is how the network performs in different formats. When deployed in TFJS format on a website, it often behaves strangely, detecting objects where there are none. On the other hand, in TFLite format, such failures are almost non-existent. If you access the link on your mobile phone and grant camera permission, you'll witness the neural network (in TFJS format) attempting to find objects even when there are none. [link] [comments] |
Predictive technology: computational social science
