Fragment the scene, either customizing the algorithm or changing the production mode.
Author | Xiu Song
Edit | Yu fast
Data, algorithms, and computing power are known as the troika of AI.
Among them, data is crucial to the effect of the algorithm model: AI with deep learning as the core, in order to avoid overfitting or underfitting, needs to use a large amount of data for model training, so that the model achieves better good fit, which is undoubtedly helpful for solving the scene problem.
However, training algorithms through a large amount of data is theoretically impeccable, but when AI moves towards the landing scene, it is not so easy.
Take the data problem as an example:
For fragmented scenarios, data becomes a difficult problem.
1. Many scenarios do not have the conditions for data collection, or the cost of collecting data is too high.
2, whether the data is valid, invalid data in addition to useless, but also the formation of noise interference, processing is also time-consuming.
3. The amount of data in some scenarios is not large enough, and it is difficult to achieve a large sample
Without a sufficient amount of data, it is difficult to train a good algorithm model, and it is impossible to solve the scene problem.
This is actually one of the major problems facing the digital transformation of the industry: the digital transformation with artificial intelligence as the core, when the actual scenario is landed, due to the shortcomings of the prerequisites for the use of technology, it is difficult to release the thrust of industrial transformation.
"From the perspective of cost and benefit, in some scenarios, AI is like a chicken rib, which cannot solve practical problems, let alone reduce costs and increase efficiency."
Wei Hongfeng, CEO and chief scientist of Zhongke Zhiyun, told Leifeng Network-AI Nuggets that under the trend of digital transformation, interlacing is like a mountain, and the entire market demand is fragmented, which poses two major challenges for AI: either to customize the scene; or to change the current algorithm production model.
But neither way can escape the data problem mentioned earlier.
For some closed scenarios, such as campus logistics, a fully automated process can be achieved through a scenario customization solution, but this is not suitable for fragmented open scenarios; in addition, the cost of customization is too high, and the independent development of algorithms will face a serious mismatch between input and output ratio.
Therefore, a more appropriate solution is to change the existing algorithm production model.
This is also the root cause of the popularity of small sample learning (FSL) and AutoML (automatic machine learning).
1 FSL with AutoML
Small sample learning is a branch of machine learning, born in the context of fragmented scenarios where it is difficult to obtain sufficiently valid data, and aims to train algorithmic models with less data volume or sample.
Compared with traditional machine learning, the advantage of small sample learning is that it does not require a large amount of data to support it, but it also brings problems: too little data leads to unreliable minimization of empirical risk.
The so-called empirical risk refers to the average loss of the model on the training sample set. Usually, the sample size is large enough (traditional machine learning) to minimize the empirical risk to ensure a good learning effect; conversely, the sample size is too small, and the effect of the empirical risk minimization learning is difficult to guarantee.
For example:
Suppose that the child is used as the algorithm model, the Chinese characters are the sample data, and the target task is to recognize the Chinese character "I". When a child copies "me" enough times, then the child's learning effect is better; conversely, if it is only copied a few times, then the child may or may not know the "me", and the experience risk is unreliable. (The case is not rigorous, for reference only)
At present, the industry has begun to solve the problem of small sample learning through data enhancement, model constraint space, search algorithms, etc.
"If the model is strong enough, the demand for samples doesn't have to be very large." Wei Hongfeng said that a good model can form self-training through small sample learning, thereby improving the accuracy and adaptability of the algorithm.
The algorithm model established is closely related to AutoML (automatic machine learning).
In the traditional AI algorithm development process, from business and problem definition, to data acquisition and labeling, storage management, data analysis and visualization, to model structure design, optimization... Finally, when it comes to application development, it takes about thirteen links, and its time cost and development cost are at a high level.
For enterprises that do not have the ability to develop algorithms, but have the needs of algorithm applications, this process is very "complex", and the cost is high and unbearable.
Therefore, the traditional AutoML that only focuses on algorithm selection and neural network architecture search cannot meet the actual needs, and AutoML, which covers the whole process of algorithm research and development, came into being, from the four aspects of feature engineering, model construction, hyper-parameter selection, and optimization methods to achieve automation, the advantage is that it not only reduces the cost of algorithm production, but also improves efficiency, and reduces the threshold of algorithm production.
For example, in the production process of traditional AI algorithms, it is necessary to manually label data and spend a lot of time processing data; it can improve the efficiency of data labeling through automatic labeling and manual re-inspection.
In the industry, there are already many more mature AutoML platforms, such as Foreign NatureLab (automatic feature engineering), Google Cloud Vertex AI NAS; domestic ai Prophet AutoML, the fourth paradigm, and so on.
In addition, there are some AI companies, such as Zhongke Zhiyun, which are also doing a corresponding similar platform (X-Brain).
As an AI company focusing on security governance, Zhongke Zhiyun mainly uses the X-Brain AI active learning platform to integrate small sample learning framework, multi-source fusion perception computing and other technologies to provide AI security governance services for the industry.
"At the heart of X-Brain is a framework of active learning algorithms that applies self-developed Active Learning techniques to change the passive acceptance of manually labeled samples for supervised learning."
Wei Hongfeng told AI Nuggets that the platform can actively determine whether the sample needs the participation of the algorithm engineer through the AI, and by only allowing the algorithm engineer to participate in the confirmation of some difficult samples, the human-in-the-Loop pattern is formed, the model is actively trained, and the model is automatically iterated.
As mentioned earlier, small sample learning is to solve the problem that fragmented scenarios cannot obtain a large amount of data to train the model, while AutoML is in the traditional algorithm mode, through active learning to improve algorithm productivity and liberate labor costs.
In other words, small sample learning solves data problems, and AutoML is a new way to improve the productivity of algorithms, and the two are combined with each other, or can solve the algorithm accuracy problems existing in small sample learning.
2 Impossible triangle?
"Small samples are the basis for low cost, because the sample size is small, and the training model does not require high-power hardware equipment." Wei Hongfeng said that how to make small sample learning achieve commercial precision is a huge challenge.
Because small sample learning has the problem of minimizing the unreliability of empirical risk, that is, the learning effect is uncertain, it can be commonly understood that the accuracy of the algorithm may be high or low.
"In some scenarios, the accuracy trained by small samples is difficult to achieve commercial level in the early stage. But AutoML can shorten the process from early model to commercialization. "
Wei Hongfeng introduced that after the sample is collected, X-Brain automatically labels through feature extraction, and automatically trains the model, which is evaluated by specific business personnel to see if false positives have occurred and adjust the parameters.
Based on these adjustments, the platform's automatic training mechanism puts the labeled data into retraining... Through this kind of circular training, the accuracy of the algorithm is improved.
In this process, the sample size is small, the model is not large, and the accuracy is improved through active training, so as to achieve low-cost use.
Therefore, between small samples, low cost and high precision, there is not an "impossible triangle".
Still take the previous child literacy as an example: a child only copied "I" a few times, and the result is that they may or may not know "me". If a teacher is introduced to guide and correct, then even if the number of transcriptions is small, it is more likely that they will know the Chinese character "I".
In this case, the teacher plays the role of a business person, and the child is the algorithm model. In the automatic learning process of the algorithm (child), the business personnel (teacher) need to adjust the parameters (guidance), and the learning effect is better.
The advantage is that while reducing the burden on children (reducing the number of transcriptions), learning efficiency (algorithm accuracy) is improved; from the perspective of the entire learning process, the teacher (business personnel) does not need to supervise the whole process, thereby reducing costs.
This actually changes the previous algorithm production model, bringing algorithm production into the era of "popularization" and "low cost".
3 AI blends with the scene
AI from the first half to the second half is actually a shift from spelling technology to fighting scenes: AI can only generate value if it lands in actual scenarios.
In the digital transformation of the industry, various fragmented scenarios have relatively high requirements for the accuracy of the algorithm. If the cost cannot be reduced according to the traditional algorithm production model, coupled with the weak purchasing power of small and medium-sized enterprises, the digital transformation of traditional enterprises will inevitably be difficult to produce.
In diversified scenarios, it is also difficult to use a general algorithm to "eat more than one bite".
"Different scenarios require different data samples, the trained models are different, the general algorithm model is not applicable, and the accuracy will be greatly reduced."
Wei Hongfeng said that small sample learning and AutoML solve algorithm production problems from the technical point of view, but how to make technology better serve enterprises and solve practical scenario problems requires that after the specific scene pain points are disassembled, they need to be integrated into the entire algorithm production process, starting from business and problem definition, to model tuning, and finally to algorithm delivery, we must "let people who understand business participate" .
This is mainly reflected in two aspects: one is to let enterprises experiment on the platform, that is, business trial and error; the other is to strengthen interaction with enterprises, so that people who understand business can participate in algorithm training.
"Customers don't pursue 100% accuracy and can accept false positives, but they can't accept that the cost of trial and error is too high, resulting in an increase in total cost." Wei Hongfeng believes that the second half of AI is driven by scenarios, in fact, it is also customer-driven, and the core is to solve the problems encountered by customers (enterprises) in the production process. The best way for AI companies to do a good job is to interact directly with business people who are familiar with the scene, rather than the AI company itself to understand the industry, otherwise the cost will be high.
For enterprises, considering cost benefits is the starting point for their purchase of technical services, and cost (including trial and error, time, manpower and other aspects) is the first element, followed by considering the benefits brought by technology.
Wei Hongfeng said that after enterprises purchase technical services, only when they achieve practical results in reducing costs and increasing efficiency, the repurchase rate will increase, and AI companies can form a positive business closed loop.
"AI can only be driven by business, not capital, to get out of the predicament and change the industry." The first premise for doing a good business is to combine scenarios.
Whether it is small sample learning or machine learning, it is only the "art" of production algorithms, and it is the "kangzhuang road" of AI to solve the pain points in the process of industrial digital transformation in combination with scenarios.
