The essence of intelligence lies in the ability to learn.

The idea of bionics:

Parallel computation, swarm computation, based on the interaction of individuals, using their existing competitive or cooperative behavior, can finally have the emergence of new properties at a specific level, like the accumulation of infinitesimal quantities of calculus, intelligence is the original function. We can understand it as a kind of Turing machine, not based on rules like mathematical analysis, but statistical analysis of probability distribution: individuals have information interaction between them, and can change their own state according to the information, which is distributed, and there is no central control center, just like the ising model/cellular automata, which can manifest specific complex patterns, and we can see it as a kind of intelligence. The emergence of intelligence is inevitable (binomial distributed). Our social mechanism is also a kind of intelligence, and there are also certain algorithms. Therefore, we can model through the accumulation of data and continuously approximate this intelligent algorithm.

The ant colony algorithm, a path search algorithm that simulates ant foraging, first iterates through all the possibilities, that is, the ants move and secrete pheromones, and then continuously optimizes according to their respective indicators, that is, other ants preferentially move in these paths and further secrete pheromones, converging to a certain optimal solution (pheromones will naturally decay, and the pheromone concentration of the shortest path is relatively high), which is an application of the lowest energy theorem. Moreover, a small number of other strategies are adopted so that the optimal solution does not stay at the local optimal solution. This is an application of game theory.

The swarm intelligence algorithm solves the NP difficult problem (combinatorial explosion, all the possibilities are too many, it is impossible to solve it in the second time of the polynomial of n, and the problem size increases exponentially), which is essentially like a Fourier transform, which is to solve the problem from different angles. It can be understood in terms of the preferential connection principle of the network, and finally produces a certain power-law distribution, so that a certain network structure can be generated, which can correspond to a certain input and produce a specific output. In fact, it is to use different information to prune, greatly reduce the scope of the search, and search in a range that is more likely to search for the optimal solution. We are convinced that there is also a power-law distribution in these ranges, i.e., a small number of regions are of greater importance. The idea of evolution, as is genetic algorithms, which are able to find the optimal solution by iteratively updating the parameters. The implementation of the algorithm needs to take into account the determination of different indicators, such as fitness, etc., as well as updated formulas, i.e., various strategies.

The mathematical principle behind it is the same as deep learning, the setting of various parameters is actually like a blind man touching an elephant to depict the different levels of a macroscopic object, and finally the combination is like calculus to achieve this high-dimensional structure.

Computational intelligence, the intersection of information technology and biological sciences, such as neural network algorithms, evolutionary computing and artificial life, etc. Algorithms that have evolved in nature are used to computation, such as the computation and evolution of the brain. The idea of machine learning algorithms.

Neural networks: parallel distributed processing, nonlinear mapping, able to learn through training, adapt and integrate, hardware implementation

Basic element neuron f(X) = ∑wixi + ?j

Neural networks: 1 recursive feedback network (neurons are connected to each other) 2 feedforward multi-layer network (hierarchical structure, no connections between the same layers)

Evolution: Variation-Selection-Survival of the Fittest. The basic motivation is to strive for a greater proportion of existence in genes and so on (selfish genes)

Genetic Algorithm: 1 Encoding and Decoding (Transforming the Problem Structure into a Bit String Form, which is Encoding, and This Encoding Form is a Chromosome) 2 Fitness Functions (Objective Functions, Metrics) 3 Genetic Manipulations (Selection, Crossing, Variation)

Artificial life, the definition of life is very broad, I think a function is a life. To understand the function of life with object-oriented thinking, we can use the idea of divide and conquer to decompose into different systems, continue to decompose into more detailed levels, down to the basic modules, and then construct the complex function of life from bottom to top.

Problem solving and searching, first need to abstract the problem into a certain mathematical model, and then find the answer to the problem, there are many means, search is one of the main means. Theoretically, there will be a corresponding answer to the problem, and we need to construct it and let the machine solve it automatically.

The state space approach abstracts the problem as a transfer of states, and then finds the sequence of states that can achieve the optimal solution through search algorithms such as depth/breadth-first search, that is, the set of operations between the initial state and the target state. Status: Where the problem is at a certain point in time, the situation, etc. These factors can be represented in the form of vectors and can be considered as features. An operator (operator, operator) is an operation that causes a state to change. The solution process is transformed into a problem of searching for a path from the initial state to the target state in the state space diagram. Modeling with graph theory: 1. Depth/breadth-first search without information; 2 There is information search, A algorithm and A* algorithm. Heuristic search, using the empirical information related to the problem, introduces the valuation function to estimate the probability that the node is located in the solution path, and then preferentially selects the path with the small value of the estimation function, and then sorts.

Search optimization ideas:

1. According to the state change from the current optimal solution to the next optimal solution, the local optimal is used to approximate the overall optimum.

2. Priority connection mechanism of the network. Depth/breadth-first search.

3. Memory of intermediate results, reducing double counting.

Pattern recognition, the diagnosis of various diseases is actually a kind of classification, through various symptoms, laboratory tests, imaging examinations, etc., the data is mapped to the list of possible diseases, and the results are sorted according to the probability of different diseases. At the beginning, it was a matter of searching, according to the database matching information (one disease corresponds to multiple symptoms, one symptom exists in multiple diseases, and this relationship can be quantified according to statistics, such as a cold has a xx probability of fever, yy probability of runny nose...). Fever has an AA probability of a tumor, BB probability is an infection...), a certain algorithm is used to construct a mapping between the symptoms and the diagnosis of a new patient. Bayesian inference may be used, such as a combination of symptoms that can be diagnosed with more certainty, as is a change in posterior probability. With enough evidence, a diagnosis with a high probability can be obtained, and medical treatment can be formulated based on this. Diagnosis, which can be abstracted as a matter of classification. The specific mechanism can refer to the machine learning algorithm, which is the black box of the whole technology. The doctor's learning and algorithm training are consistent and are typical of supervised learning, learning from labeled data, which enables new inputs to be mapped to specific outputs, i.e. classification.

Existential hypothesis: As long as it is a disease, there is a cause. Such as microbial infections. We believe that even if there is no very clear cause, it can be found through the structure, such as environment + genetics. Considering different influencing factors as states, understood by the Markov model, the transfer of these states corresponds to the symptoms of different diseases, that is, the combination of different factors may be the etiology at the abstract level.

The classification of diseases begins with the modeling of the disease, which is the result of artificial definitions.

Pathological diagnosis is a morphological reflection of the body's condition, which can have a greater strength of evidence to make a diagnosis. We can classify according to multiple indicators, such as various disease markers, positive cells/molecules, cell morphology, etc., as long as there are enough indicators/parameters, Hilbert infinite dimensional space, theoretically can model and classify all diseases. Of course, it is sufficient to take a limited number of metrics of greater importance, as is the case with PCA. These key indicators can be regarded as the basis of linear algebra, and the whole space can be constructed through linear combination, as long as its error from the real situation is less than a certain degree, we regard it as equivalence, which is the idea of analytical mathematics. Of course, these key indicators are the results of statistics.

The treatment of the disease can be understood as the choice of strategy, symptomatic treatment is the treatment of the relevant refinement structure of the target, such as cooling the fever and relieving the pain, theoretically adjusting all the symptoms to the normal situation is the treatment, but in practice it is impossible to take so many measures (limited resources), we can only deal with the objects of greater importance. Further, it is the treatment of causes, abstracted to a higher level, like the level of the original function of calculus. Only by finding the optimal solution can the best treatment be given to the patient, but the reality is often to act at the opportunistic level, hoping to approach the overall optimal with the local optimum, because the doctor's attention resources are limited, it is difficult to get the optimal solution immediately, and can only follow the experience to deal with it quickly. Such as various clinical guidelines, expert experience, the results of various clinical trials. It is a strategy to make the best decisions with limited resources and it is impossible to be perfect.