Neuroscience: Mind Reading (Neural Decoding)

Early research Over the years many companies have studied this technology. In 2007, Microsoft wanted to evaluate how people interact with computers. They began researching the use of the electroencephalography (EEG) to record electrical signals when users use the computer. Microsoft is still researching this problem. In 2009, Toyota announced the idea of ​​"Mind Reading" wheelchair. This chair is designed to help people with disabilities and the elderly more freely move around and become more active people.

"Mental reading" AI: Purdue Engineering Wen et al. Perform neural coding and decoding using deep learning of dynamic natural vision. Researchers showed how to decode what the human brain can see by interpreting fMRI scans of people watching videos using artificial intelligence. YouTube video is here. The original text is here. Dynamic routing between capsules is described in Hinton et al. A new paper by Jeffrey Sinton. A capsule is a group of neurons whose activity vectors represent the instantiation parameters of a particular type of entity, such as an object or part of an object. We show that a differentially trained multilayer capsule system achieves the most advanced performance on MNIST and is far superior in identifying numbers that are significantly overlapping than convolution networks. PyTorch implemented the paper here

The proposed neural network architecture called RNN Encoder - Decoder consists of two recurrent neural networks (RNN) which act as a pair of encoder and decoder. The encoder maps the variable length source sequence to a fixed length vector, and the decoder re-maps the vector representation to the variable length target sequence. Two network collaborative training to maximize the conditional probability of the target sequence for a given source sequence

Technical report on ENet is here. ENet is a network of encoders and decoders. The encoder is a conventional CNN design for classification and the decoder is an upsampling network for propagating categories to the original image size for segmentation. This method uses only neural networks and there are no other algorithms for performing image segmentation. ENet is designed to use as few resources as possible from the beginning. Therefore, a very small footprint is realized, in which both the encoder and decoder network occupy only 0.7 MB with fp16 precision. Even with such a small size, ENet is equal to or better than other pure neural network solutions in terms of segmentation accuracy.