- Startup in China uses gait recognition technology to identify people. Said to be accurate even with disguise and faked limps. Currently it is tested for criminal investigations, but it screams intrusion, especially in CN.
- Deep learning optimizes based on parameters and weights of a neural network through a stochastic gradient descent (backpropagation) which can be costly an inefficient. “CoDeepNEAT” assembles layers of neurons in an artificial neural network by mixing different functions in each layer, with convolutions of a CNN or a ‘cell’ of a recurrent neural network.
- To meet the challenge of AI for enterprise, IBM draws upon the capability maturity model. Fundamentally, AI deals with probabilistic than deterministic models, and requires a lot of “messy data”. IBM’s paper calls for AI to be customized to its business case. However, ground truth labels on data detracts from unsupervised machine learning. A Training Lead advocated to lead feature extraction contradicts automated feature extraction in deep learning. Lastly, there are no clear guidelines on how enterprises can keep their models unbiased.
- Machine learning with images and videos require a richly annotated / labeled data set. Researchers at Sun Yat-Sen use a neural network model to continuously compare guesses of multiple networks with one another, reducing the need for ‘ground truth’ from a labeled dataset. A convolutional neural network (CNN) extracts the 2D stick figure before a long short-term memory (LSTM) neural network specializing in retaining memory of a sequence of events extracts continuity of the body from multiple video frames to create a 3D model. The learning phase is self-supervised, correcting its comparison between 2D and 3D models.
- Comment
- Reblog
-
Subscribe
Subscribed
Already have a WordPress.com account? Log in now.
Q.bit 