No-Code, Low Code Machine Learning Platforms Still Need People

dWeb.News Article from Daniel Webster dWeb.News

No code, low-code (horizontal), machine learning platforms can be useful in scaling data science within an enterprise. Despite this, many companies are discovering that there are many ways data science can fail to solve new problems. Zillow suffered billions in losses due to its flawed data-driven home value model. Facial recognition software has been shown to bias hiring decisions towards protected classes when it is used as a data-driven human resource technology.

Automation is an excellent tool, but you must consider the limitations before using a horizontal ML platform. To be reliable and add value over time, these platforms must be configurable, flexible, and easily monitored. They must allow data to be weighted in flexible, user-controlled ways. Data visualization tools are also required to identify outliers and contribute to noise. To alert users of changes, they also require automated model parameters and data drift monitoring. We haven’t advanced beyond the point that algorithms can outmatch human intelligence, as you can see.

Don’t let AI/ML/low codes fool you… people are still needed. Let’s look closer at the reasons.

Machines Learn from Humans

Trying to replace domain experts, data scientists, and engineers by automation is a risky proposition that could result in disaster when applied to mission-critical decision making systems. Why? Human beings can understand data in ways that computers cannot.

Humans are able to distinguish between data errors (e.g. Game/Stop/GME trading in February ) to align unusual data patterns and real-world events (e.g. 9/11, COVID, financial crises, elections). We also know the impact of holidays and calendar events. Depending on what data is used in ML algorithms, and the data being predicted the semantics of that data may be difficult for automated learning algorithms. If these relationships are not hidden from the operator, forcing them to discover them is unnecessary.

The most difficult part of data science aside from semantics is distinguishing between statistically sound results and useful results. Although it is easy to believe that you have better results using estimation statistics than you do with the actual model, neither model is more useful for solving real-world problems. Even with statistical methods that are valid, there’s still an element to modeling results that requires human intelligence.

When developing a model you will often face issues with the estimation statistics. These include how to weigh them, evaluate them over time and determine which results are significant. You also have to consider the issue of “learn”, which is when you test too often on the same data set. This can lead to unrealistic test results. You must also build models and determine how to combine all the statistics into a simulation method that is practical in real life. It is important to remember that just because a machine-learning platform was successful in solving a particular modeling and prediction problem does not mean that it will work on another problem in the domain or vertical.

There are many decisions that must be made during the data science research, design, and deployment phases. Experienced data scientists are needed to design experiments. Domain experts are required to understand the boundary conditions and nuances of data. Production engineers must be able to explain how models will be used in real life.

Visualization is a Data Science Gem

Data scientists can also benefit from visualizing data. Visualizing data is a manual task and more art than science. The ability to plot raw data, predict the correlations between them and their quantities, and create time-series coefficients from these estimations can provide valuable observations that can be used in model construction.

You might notice a pattern in data or anomalous behavior around holidays. Extreme moves in coefficients could indicate that your learning algorithms are not properly handling outlier data. It is possible to notice differences in behavior among subsets. This could indicate that your learning algorithms are not able to handle your data well. Self-organizing learning algorithms are another option to uncover hidden patterns in data. A human being may be more equipped to identify these patterns and then use the insights to improve the model building process.

Horizontal ML Platforms Need Monitoring

Model monitoring is another important role that people can play in the deployment and maintenance of ML-based AI systems. The type of model used, the predictions it makes, and the production process are all important factors in monitoring the model. This allows for deviations to be tracked and can be prevented from causing real-world problems.

If models are being retrained regularly using more recent data it is important that you track the consistency between the new data and the data used previously. It is crucial to ensure that new models, which are trained using more recent data, are used in production tools. This is especially true if the models are model- and task-dependent.

While automation can be a great tool for solving many problems in many industries, human intelligence is still essential to these advancements. Automating human behavior can be done to a certain extent and in controlled environments you can replicate their power and performance using ML-based AI systems that are low-code and no-code. In a world that machines are heavily dependent on humans, don’t forget about the power of people.

Read More

For more dWeb.News Technology News

The post No-Code, Low Code Machine Learning Platforms Still Need People appeared first on dWeb.News dWeb.News from Daniel Webster Publisher dWeb.News – dWeb Local Tech News and Business News

dWeb.NewsRead More

Similar Posts