Talking nonsense, finding the rhythm--what is the future of machine learning

Template Matching

What is the most straightforward and straightforward algorithm to describe something like--KNN

What is the template matching algorithm--kmean to learn the most close to human cognition? Kmedoid? or other clustering methods?

--it could also be encoded with a simple convolution kernel.

What structure is closest to human cognition

Generation model Evolution-generation + discriminant model

Discriminant model should be just one kind of cognition, not the whole of cognition, the whole of cognition should be conceptual, but the discriminant model can't embody it.

But it is not that a model with a concept can help to classify the discriminant model well.

-the key to the above problem is the dependence of the discriminant model on the data characteristics.

--discriminant model depends on the richness of the data, this richness should be as small as possible in-class divergence, the amount of each sub-class is large

Gradient Descent learning algorithm or the whole DNN model is quite like a black box, do not know the law of the change of weight, or is not a gradient descent algorithm to adjust the weight value-this point by the input of mnist Softmax_loss_layer can be seen, Fully connected the features that CNN extracts seem to be disrupted, and must be "normalized" with Softmax to make the data look like a coded result.

Do not think that in recent years can be a large-scale attack on artificial intelligence, artificial intelligence there are many ways to go, even to solve the human brain cognition, visual characteristics, the characteristics of other organs are worthy of research direction, there is a long long way to go

The development History of machine learning:

Lack of data, lack of theory and theory, lack of data (the rise of shallow models), no lack of data, no lack of theory, lack of acceleration (DNN Eve), no lack of data, no lack of theory, no shortage of accelerated (deep learning), a small number of label data, A large number of unlabel data, lack of reasonable "cognitive structure"--from brain cognition to the cognition of each organ

Semi-supervised model

The idea of hyper-categorization is still not pretty, a bit like a solution to the problem, but not necessarily

The sensory generation model is not only used in parallel with discriminant model, but also in serial

Talking nonsense, finding the rhythm--what is the future of machine learning