What I learned about machine learning after working in 12 startups

Hello.

Having worked in 12 machine learning startups, I made eight useful insights about products, data, and people.

All startups were from different areas (fintech, biotechnology, healthcare, learning technologies) and at different stages: at the pre-seed stage, and at the stage of acquisition by a large company. My role has changed. I was a strategic consultant, the head of the data analysis department, filled up by a staff member. All these companies tried to create a good product, and many succeeded.
')
During the work I came to the following conclusions:

Product is more important than AI

These startups are developing products, not learning artificial intelligence. As a convinced mathematician, at first I was more interested in machine learning and the creation of new methods and algorithms.

I soon realized that even accurate machine learning models are not valuable in their own right. The value of AI and machine learning directly depends on the value of the product in which they are used. The goal of a startup is to learn how to create machine-based products.

With this approach, it sometimes turns out that machine learning is not the most effective tool. Sometimes it's not the task, but in the process of solving. Even in such situations it is useful to turn to scientists: they use a scientific, data-based approach. However, do not waste time on AI where you need to correct the process.

Aim for synergy between data and product.

You cannot create something valuable by adding predictions based on a machine learning model to an existing product. Strong AI is not an addition to the product, it is the basis. In such cases, it is the AI ​​that creates the value. Such products are developed taking into account this fact: in them both the product and the data work in synergy.

A good performance results in an interaction, which I call “a combination of product and data.” The product fully realizes the potential of the data and at the same time generates new data necessary for improvement.

When working on AI, you need more than just engineers and data scientists. Work on the value of the product goes faster if other members of the team participate in the discussion, from product managers to managers. This requires a level of knowledge and involvement to which even engineers working in start-ups are not used to.

First the data, then the AI

For AI and machine learning you need a lot of high-quality data. When creating a product from scratch, think about data collection from day one. Before introducing artificial intelligence technology into an existing product, get ready to invest a lot in data engineering and architecture changes.

First, find out the value of the product, and only then proceed to work. The better the data processing, the more informative the analytics is - it is crucial for the development of the company. So you will demonstrate the value of the product and attract investors. Start thinking about machine intelligence when analytics is reliable.

Invest in communication

To create a product, you need qualified product managers and management support. Strong AI and in-depth training are of interest to many, but people who are far from the IT industry do not understand these technologies. To discuss machine learning and AI, you need to understand the statistics: ineffective communication leads to unrealistic expectations.

The product manager and data engineers must constantly discuss business metrics and their transformation into a product. This is especially important for engineers: in order to work effectively, they need to deepen their knowledge in their field and in business.

"Simple and obvious solutions" are not so obvious.

As I mentioned above, it is often easier to solve a given problem using simple and obvious methods. This is partly because today's “simple and obvious” solutions yesterday were complex and original. Now using word2vec is as simple as regression . Every day there are more new tools, and understanding of these tools is important for the data analyst.

The emergence of open source tools has led to the fact that now proprietary platforms in machine learning is not an effective solution. Of course, you should use proprietary algorithms if they are effective in your industry and to solve your problem. But let's leave the study of deep learning staff Google - focus on business problems.

If in doubt, show the data to users.

At an early stage, it is important to establish a feedback with the market. However, machine learning requires data that takes a long time to collect. This is the problem: how to analyze a picture without large amounts of data?

Most often, the best solution is to show the accumulated data to users. It doesn't matter that you have little data: people only process small amounts of data at a time. See how users interact with data: what do they ignore, and what do they want to understand more? So you will realize how potentially valuable your business data is.

Build trust

Trust is the basis of the success of most technologies: people want to trust the technologies that they use. Some people worry that automation will take away their jobs, others rely on technology to make important decisions. In both cases, trust in applications and machine learning algorithms is important.

If artificial intelligence does not help a person to make decisions, but decides instead of him, users quickly lose confidence in the application.

Trust is easy to lose and extremely difficult to regain. Create products that people will trust.

Share the article with colleagues, draw conclusions and work productively. If you have something you can share - write in the comments. More information about machine learning in the Neuron telegram channel (@neurondata).

All knowledge!