k-means learning: learning unsupervised dictionary

[Eugenio Culurciello]

We heard a lot about Hebbian learning in the brain. "Neurons that fier together, wire together". But as of today, there is no convincing example of large-scale brain-like learning that can outperform the state-of-the-art machine learning algorithms (not counting living things...)

Now new exciting deep learning techniques that are extremely simple allow us to tie brain-like learning to simple machine learning code that can learn dictionaries of unsupervised templates in minutes, and with no parameters!

This technique, currently named "k-means learning" uses clustering (neurons that fire together) to combine neural outputs (neurons wire together) to form base functions of the input data. The beauty of this technique is that it uses simple averages of input data to generate the base functions, thus it might just be the simplest deep learning method out there!

Stay tuned: more info and data to come. 

[Eugenio Culurciello]

the figure represents the first layer of base functions we extracted with this technique

exciting!