Outline
1. Neural Network Learning
2. Common Model
3. iOS CoreML like in closed source library
4. DNN network implemented using Sigmoid activation function and by bench_ios ReLU in CC_ANN.
5. caffe, tensorflow iso-contrast
https://baijiahao.baidu.com/s?id=1574518061092294
Why can people think? Scientists have discovered that because of the body's neural network.
Nerve endings by external stimuli, into electrical signals transduced into nerve cells (called neurons). Numerous nerve center of neurons. The nerve center of the comprehensive variety of signals, to make judgments. According to the human nerve center of command, respond to external stimuli.
http://www.21ic.com/app/computer/201707/730987.htm
special case of exclusion
https://baike.baidu.com/item/%E5%9B%BE%E7%81%B5%E6%B5%8B%E8%AF%95/1701255?fr=aladdin
memory capacity
Q: Can you play international Chess it?
A: Yes.
Q: Can you play chess you?
A: Yes, I have not told you yet?
Q: Please answer again, do you play chess?
A: You do not trouble trouble, why put the same old problem.
https://m.huxiu.com/article/247451.html?from=singlemessage
man-machine dialogue
http://www.xinhuanet.com/science/2017-03/07/c_136108535.htm
For example, "one plus one equals a few" and "Xiao Ming have an apple and a pear, asked Xiao Ming has several fruits" the same is essentially a "1 + 1 =?" the two questions on the kinds of questions the concept is the same, but different expressions. The computer needs to know how to put the above two issues are abstracted into two objects are added, which involves the so-called natural language understanding.
Hosseni 2014 trained classifier determines a verb belonging to a plus / minus
https://www.leiphone.com/news/201605/sQdUMNhFoWhoATv8.html
looking SVO
https://www.2cto.com/kf/201608/534598.html
the HMM Markov assumptions: a word appears depends only on a limited or a few words before it appears. If there is a word depends only on the front of a word it appears, then we call it bigram. I.e., P (T) = P (W1W2W3 ... Wn) = P (W1) P (W2 | W1) P (W3 | W1W2) ... P (Wn | W1W2 ... Wn-1) ≈ P (W1) P (W2 | W1 ) P (W3 | W2) ... P (Wn | Wn-1)
MIT Open Course: Artificial Intelligence
http://open.163.com/movie/2017/9/Q/S/MCTMNN3UI_MCTMNR8QS.html
What leaves from the tree to find the book
You will run the streets carrying a bucket of water in your shoes wet in the network can not be found to tell you that no one but you can imagine
Goal tree
And/or tree
Goal tree how/why 向上 向下 60年代西蒙
heuristic transformation
Composation depth
身边的同学能解出一个数学题 你解不出 你觉得他很智慧 当你做了这个题100遍 看到它就会解时 你就觉得他没有那么智慧 他解题方法和我一样 我也会
Complexity of environment not complexity of program 蚂蚁回家绕来绕去走到家门口
行为复杂度是以上的较大值
Forward chaining based rules “expert” system
Backward
Deduction system /prove get new
不能预测特例
What is this? 正向
Is this a xxx? 反向
Search is about choice
Depth first search 遍历每一条到底直到找到/ breadth first search / hill climbing(sort,keep W best) / beam search
除非节点没有到过（避免重复）除非没有扩展过/ doesn’t work when we add admissible heuristic
Backtracking idea回溯
Sort is expensive
Admissible (H<D) Consistence(|H-G|<D)
在某些部分很像 那么他们在其他方面可能也很像 学习先例 case 的意义
Careful about confusion of correlation with cause
关联性 因果性 猫和狗看到人和节食饮料 不会想到是为了减肥
They see the correlation, but they don’t understand the cause.
Identification tree; decision tree
找界限值 阈值
Neuron
Dendritic tree
第一次计算赋值w1，w2，反向修正w2，到w1
计算离结果最近的w，一直往回走，局部计算、反向传播法back propgation; liner
Overfitting 拟合过度
回到最前，是新的一代，一部分被淘汰
有丝分裂 mitosis 染色体 chromosomes
有一些突变因子导致概率变化，再使用交叉互换
P1被选中的概率=P P2被选中的概率等于剩下的（1-P）乘以P2被选中的概率
会被困在局部最大值
多样性
第一个考虑适应性，没有选择，之后先考虑多样性，选哪个和已经选的差异最大，这就得到最高多样性排序
得到最优解步骤：先发散后收缩到最优
Dogs 和 cats s的发音不同 phonological rules
plural
Short-Learning One
Initial Model
Papers famous five elements
Symbol
Slogan
Salient prominent idea is a one-time learning can be achieved by the use of similar errors
Surprise
Story
To understand the meaning, there is a change table, each phase change what
AI mechanisms to improve writing, so that the reader can easily understand, reduce the burden of understanding
1, do not use pronouns
2, do not use the "former", "the latter"
3, Don ' t call a shovel a spade with synonyms not so cut to cut
Genel problem solve
SOAR / state operate and result
Instinctive reaction, reaction learning, thinking reactive, reflective thinking, self-reflection, self-conscious thinking
iRobot robot vacuum cleaner
https://baike.baidu.com/item/iRobot%E6%9C%BA%E5%99%A8%E4% BA% BA% E5% 90% B8% E5% B0% 98% E5% 99% A8 / 10537215? fr = aladdin
How to become smarter?
Fast talking will hinder your language processor, no time to think, let you get off to buy, buy him something
Probabilistic reasoning I