I have worked with 12 startups. They range from financial technology and healthcare to the vertical areas of educational technology and biotechnology, from seed to acquisition. My responsibilities vary from the first employee to the data science director and strategic consultant. In all of this, I am committed to interesting machine learning and data science issues. Everyone is trying to make quality products. Many people have succeeded.
This is what I learned.
This is about building products, not about AI.
As a card mathematician, I was the first to be motivated by the challenges of machine learning science and the introduction of innovative new algorithms and methods.
But I quickly realized that even the most accurate machine learning model cannot create value on its own.The value of machine learning and AI lies in Measured according to the products it provides.Figuring out how to do this effectively is the real purpose of building ML-driven products.
This is about the problem, not about the method.
If the goal is to make a product, then machine learning and AI are the means to an end. The important thing is how well they solve your product problem, not which method you use. In most cases, a fast and dirty solution will put you on the right track. Don't train deep neural networks when simple regression is all right.
When you focus on the problem, sometimes you may find that machine learning is not the right tool to solve the problem. As it turns out, many problems are mainly related to the process. Even in this case, data scientists naturally tend to adopt a rigorous, data-driven approach, so they can contribute a lot of value. But this does not make it a good idea to use AI to fix bad processes. Fix the process.
Find synergies between data and products
The true value of machine learning rarely comes from the use of existing products and the combination of predictions from machine learning models. Of course, this will add some incremental values. But in a powerful AI product, machine learning is more than just an add-on. It is the engine of value creation, and the product is built with the engine in mind: the product and data must work together.
If done well, this will lead to onePowerful virtuous cycleI call it " Product/data fit ": The product effectively realizes the potential value of the data while continuing to generate the necessary data to further improve the product.
In particular, AI can't just stay in the data science and engineering teams. From the product to the executive, other parts of the organization need to engage in dialogue to accelerate the value creation process. This requires a lot of education and investment, even beyond the capabilities that engineers are often accustomed to from building software (even in startups).
Data first, artificial intelligence first
Machine learning and AI require a lot of data and, more importantly, high quality data. If you want to build a product from scratch, pleaseConsider from the first dayStartData collection. If you want to introduce AI technology into your existing products, you are ready to invest a lot of money in data engineering and re-architecting before entering the AI section.
This does not mean that you have to do all the work before you realize any value. Better data operations mean better analysis, which is critical for any organizational learning and improvement. Use these wins to demonstrate value and generate organizational identity. And, when your analytical work is even more powerful, you can start thinking about real machine learning.
Invest in effective communication
Building great products requires the support of outstanding product managers and executives. Although many people are attracted by the power of AI and deep learning, few non-technical people really understand these technologies. An effective discussion of machine learning and AI requires a thorough understanding of the statistics, resulting inCommunication gaps, which often lead to unrealistic expectations.
A key element is to maintain an ongoing dialogue about business metrics and how they translate into modeled metrics. This puts a lot of responsibility on the product manager, but data scientists have to shoulder the same responsibilities, they must develop domain expertise and have a deep understanding of business considerations to make a real difference.
Fast and dirty is actually not so dirty
As I mentioned above, the quick and dirty method will benefit you a lot. Part of the reason is because today's fast and dirty is slow and precise yesterday. Such asword2vecThe tool has become almost as easy to use as regression, andContinuously launchedA powerful new tool. For any data scientist, a deep understanding of the various building blocks and the bonding between them is critical.
As a result of the explosive growth of open source tools, in most cases,Developing a proprietary ML platform is not a good idea. Of course, you should have a proprietaryalgorithm,These onesalgorithmUse well-known building blocks and adapt them to your problem and your field. But leave the deep learning research to Google's people - focus on business issues, remember?
If in doubt, please show the data
The most important activity in early product development is to get market feedback. But machine learning requires a lot of data and it takes a long time to get it. This raises the question: How do you gain market insight into data products without large amounts of data?
The best solution is usually to simply the userDisplay Data. Humans can only process a small amount of data at a time, so it doesn't matter if you don't have much data. How do your users handle the data you display to them? Where do they cover up and where do they want to dig deeper? Disclosing information that was previously inaccessible can be very useful and can give you powerful guidance about the potential business value of your data.
Trust is a major factor in the success of most technologies. Ultimately, every technology is used by humans and must be trusted by humans. In the context of machine learning applications, some of them may be concerned that their work is being automated. Others rely on the information provided by your technology to make important decisions.
AI products such as this (for example by tryingforHumans make decisions rather than empowering people to make decisions that will exacerbate these concerns, which will lead to a rapid decline in trust.
Trust is easy to lose and difficult to recover. Create products that people trust.
This article was transferred from awardsdatascience,Original address