Not every problem can be solved through machine learning, and not every enterprise is ready for AI applications.
For example, companies need to determine specific application scenarios, whether there is enough data to analyze, to build predictive models, to have people and tools to define models and training models, and so on.
To this end, this article will elaborate on the 10 preparations that companies need to do when using artificial intelligence, deep learning, and machine learning.
-1- has enough data
Sufficient correlation data is a necessary condition for prediction and feature recognition. So, how much data does the company need? Whether you are doing common statistical predictions, machine learning or deep learning, the more factors you have to consider, the more data you need. In general, machine learning requires more data than statistical prediction, and the data required for deep learning is more than double.
Taking sales forecasting as an example, due to the strong seasonality of the retail industry, companies are required to accumulate statistically significant monthly data for many years so that they can correct monthly cyclical changes and establish annual trends for standard use. Time series analysis model.
For example, companies can use statistical models to analyze the monthly shirt sales of nationwide chain stores over a five-year period, and based on this, predict the total sales of shirts for the next month, and the percentage of sales of women’s shirts such as a certain percentage of total national sales. The sales of blue short-sleeved shirts account for a more specific number of percentages of total shirt sales. Of course, in this process, companies also need to pay special attention to the gap between actual results and model pres.
And if you want to consider external factors, such as weather and fashion trends. Enterprises can also introduce historical meteorological data into the model for testing. Of course, doing this in a time series statistical model can be difficult, but you can try to implement it using a decision tree regression model.
As shown above, companies can try out the other seven machine learning models for regression, and then compare the “cost†(normalized error function) of each model with the actual results of last year to find the best model.
For example, when companies want to know if navy blue clothes will sell better or worse next year than they did last year? You can view all monthly sales of Navy Blue apparel and predict annual fashion trends, or you can introduce them into a machine learning model for analysis. In addition, companies can manually correct models based on information obtained on fashion media.
In this process, by creating a deep neural network, you can make predictions better. You may even find that for each hidden layer you add, you can increase the regression error by a few percentage points until you increase the next layer and you can't improve the effect. Among them, the key to diminishing returns may be because there are no more functions in the model to identify, or it may be because there is not enough data to support more room for improvement.
-2- Data Scientist
Of course, companies must have a person who can build all the models mentioned above. He needs sufficient experience, intuition, programming skills and a statistical background.
Even today's many machine learning products and service providers say that "anyone" or "any business role" can use their pre-trained application machine learning model. But the reality is that enterprise data may not be applicable to existing models. Therefore, data analysts and data scientists are needed to guide and help train the model.
-3- Track and collect important relevant variable data
In addition, companies need data with relevant variables to collect and use data from multiple dimensions to avoid statistically stated “unknown variancesâ€.
Of course, the measurement and collection of certain independent variables, such as weather observations, is difficult, impractical, or costly. In the chemical field, for example, when you plate lead to copper, you can measure the temperature and concentration of the fluoroboric acid plating solution and record the voltage on the positive electrode. However, this requires that the peptide contained in the solution be in an appropriate amount to obtain good adhesion. Otherwise, you can't know the amount of this key catalyst, and you can't use other variables to explain the change in electrode plate quality.
-4- Find ways to clean and convert data
Often, the data is very complex and its quality is often uneven. For example, during the acquisition process, one or more values ​​may be missing, individual values ​​may be out of scope or inconsistent with other values, and the person answering the question may not understand the problem or make an answer.
This, on the other hand, means that the data filtering in the analysis process takes the most effort, which may even account for 80% to 90% of the total analysis time. If all data is kept in the data warehouse or data lake during ETL (extract, transform, and load), it is possible to save irrelevant or low-quality data.
Of course, even accurately filtered data may require further conversion to perform a good analysis. Similar to the statistical method, the machine learning model works best when the parameters of each possible state are similar, that is, when the ranges of all variables are normalized. Therefore, companies must find ways to better clean and transform data.
-5- Revisit all variable data and its relevance
Next, we need to take a step back and look at all the variables and their dependencies.
Exploratory data analysis can quickly display the range and distribution of all variables, regardless of whether the relevant variables are interdependent or independent, where the cluster is, and where there may be outliers. When a company has highly relevant variables, it is often useful to remove one or more variables from the analysis. Companies can also perform methods similar to stepwise multiple linear regression to determine the optimal variable selection.
However, this does not mean that the final model is linear, but before introducing more complex factors, you need to try a simple linear model; if there are too many technical terms in the enterprise model, the final result will be a A system model determined by a variety of factors.
-6- Try to find the best model by trying again and again
There is only one way to find the best model for a given data set: that is, try all of them.
If the company's goal is to target an exploratory but challenging area (such as image feature recognition and language recognition), it may only try a so-called "best" model. However, these models are typically the most computationally intensive deep learning models, for example, with convolutional layers in the case of image recognition and long-term short-term memory (LSTM) layers for speech recognition. If companies need to train these deep neural networks, they need more computing power than the office environment.
-7- Have the computing power needed to train a deep learning model
The larger the data set, the more layers in the deep learning model and the longer it takes to train the neural network.
One way to deal with training time is to use a general purpose graphics processing unit (GPU). A K80 GPU works with the CPU, and its training speed can usually be 5 to 10 times that of using only the CPU. If the enterprise can integrate the entire "kernel" of the network into the local memory of the GPU, the training speed can even reach 100 times that of using only the CPU.
In addition to a single GPU, enterprises can set up coordinated CPU and GPU networks to solve larger problems in less time. Unless you're willing to spend a whole year training a deep learning model and have a huge budget, you'll find that renting a GPU on the cloud is the most cost-effective option. Several deep learning frameworks, including CNTK, MXNet, and TensorFlow, support parallel computing of CPUs and GPUs, and have reasonable scaling factors that can be used to support GPUs for very large virtual machine (VM) instance networks.
-8- Learn to adjust or try different methods
A simple statistical model test sets standards for the operation of a company's model through machine learning and deep learning. But if you can't use a given model to improve the level of analysis, you should adjust or try a different approach. For example, you can set up multiple model trainings in parallel under the control of the hyperparametric algorithm and use the best results. One stage.
-9- Deployment forecasting model
Ultimately, trained models can be deployed and run on servers, in the cloud, on personal computers, or on mobile phones for enterprise real-time applications. The deep learning framework provides a variety of options for embedding models into web and mobile applications. Amazon, Google, and Microsoft have also demonstrated their practices in this area, and even consumer electronics and smartphone apps that can be operated by voice recognition.
-10- Regularly update the model
Of course, you may also find that even for a well-trained model, as the data changes over time, the error rate of the model increases over time. For example, the sales model of the company will change, the competitors will change, the style will change, and the economic situation will change...
To this end, most deep learning frameworks have the option of retraining old data and replacing the original forecasting service with a new model. If you update regularly every month, you can basically keep up with the times. Otherwise, your model will eventually become too outdated and unreliable.
Bluetooth Mini Projector
Sound can be transmitted wirelessly, no audio source cable is required.
1. The Bluetooth function of the projector can be connected to a Bluetooth speaker, and you can enjoy better sound quality when watching movies and playing music;2. After the projector is connected to the mobile phone through the Bluetooth function, the projector acts as a speaker and can play music from the mobile phone.
wifi bluetooth projector,bluetooth home projector,bluetooth protable home projector
Shenzhen Happybate Trading Co.,LTD , https://www.szhappybateprojectors.com