Deep learning at the start of the 2011-year Fire (Hinton), people would think that the learning (DL) is approximately equal to convolutional neural network (CNN), a supervised learning image recognition tool;
Then came the word vector (Word2vec), people began to think that DL can also solve a part of the NLP problem
Then long short term memory (LSTM) Suddenly more successful, people began to think that DL can also do time series prediction and sequence recognition
Then DeepMind published deep Q-network (DQN), people began to think that DL can also do exercise control
Wait a minute
In short, this year, the DL gives the impression that
1, responsible for the underlying signal processing (visual, voice, text) classifier
At this point, the DL seems to not solve a lot of problems, such as?
The mass media will give the impression that the DL only has a slight advantage over the classification of high-dimensional data. However, RNN can theoretically fit all information processes, including program, logic, symbolic-based reasoning, world model deduction, sentiment, aesthetics, creativity, fine movement, and today RNN is not used to fit these procedures and processes.
I think just because we can't get a valid training set for the relevant task. We can get a lot of pictures and texts from the Internet, which make up the raw material of training neural network. But we don't share the details of all the thinking in programming and logic reasoning, and the human brain translates language symbols into intermediate information through some prior knowledge and process, and the lack of intermediate information invalidates the training set. So it's the only option to build an intelligent system from the bottom up
1. World model deduction, or common sense; When the cup moves out the edge of the table, the human brain will know that it will fall down in the next second. The training set that trains this process relies on the representation of the underlying data, and the problem with DL is the lack of a unified representation of the underlying (image) concept, which I call the DL's [underlying characterization bottleneck].
2, emotion and aesthetics, people will assume that the computer is rational, and the mood is irrelevant, however, human can express themselves through the expression, intonation, through the collection of these training sets can easily set up from the state to the mood of the map, so that the computer has emotions.
3, the essence of creativity is the fine-tuning and reorganization of data, because of genetic algorithms and simulated annealing and other optimization processes are unfamiliar, people will mistakenly think that creativity is very NB intelligent activities. The creation of the real world has two prerequisites: How to split the original data is the precondition of reorganization (recognition problem) evaluation function (evaluation function fitting problem), both of which can be solved by DL.
4. Symbolic (atomic characterization) based reasoning (instead of decision trees), such as IBM's Watson diagnostic system, still relies on many of the atomic representations of reasoning. Why hasn't this kind of thing been replaced by DL yet?
5, fine movement; the movement of each muscle of human is implicit; needs retraining, human movements rely on imitation (underlying characterization bottlenecks), reinforcement learning (underlying characterization bottlenecks), time series predictions for rewards (underlying characterization bottlenecks)
If we set up a unified representation mechanism from the underlying signals, we will form a unified concept; the DL can do the mapping between any concept combination rather than just the original image or speech classification.
What other little-known potential is there in deep learning?
