The idea of machine learning and autonomous systems has been around since the 1950s but only now has it been possible to implement them. The simple explanation resides in the huge drop in the cost per gigabyte. The distributed computational power of the cloud also solves the issue of ownership of a dedicated server to run the necessary training sessions. Now you just have to pay a moderate fee for the usage rights, instead of buying and continually upgrading an entire infrastructure.
With these logistical premises in place, there are still some boxes to tick to ensure the success of a machine learning project in your organization. Here are the top X items you have to pay attention to.
Know the limitations of machine learning
Although there is a definite buzz around Machine Learning (ML), at the end of the day it’s just another tool in a set of possibilities. It’s best suited for problems which require large amounts of unstructured data to be processed to identify trends and patterns.
You should not select this if you have a problem which requires explaining the way the algorithm got a certain result since the ML network works as a black box. Also, since building a useful model needs consistent amounts of training data and at least some resources, this is not the choice for simple, low-budget projects.
Define the problem
To decide if you should use ML for the idea you have in mind, first define the problem. State clearly what the inputs are, the expected output and draft a definition for the deliverables of the project. Ask the questions which you hope to answer and even make a hypothesis about the results you are going to get.
Ensure financial feasibility
Consider your machine learning project as a way to improve your bottom line. Most likely it will involve making some investment, and you must have a sound plan to recover the initial money put into the project, as well as generating more.
You must strive to create new revenue streams through your machine learning projects by creating items clients are ready to pay for. Think about issues they highlighted through customer service, improve their experience or gives better functionality.
Check data availability and accuracy
All machine learning projects have one thing in common. They need loads of data. Before even thinking about such a project, it helps to check if you have reliable data sources for the proposed project. Keep in mind that you will need training data, model calibration data and then a feed of regularly updated data.
Machine learning models usually work with unstructured data which, most of the times, can’t be used in the format it is retrieved. Allocate time and a budget for data preparation, formatting and cleaning.
Don’t only rely on your own data; it is better to think outside the box and combine proprietary data with widely available external data. Keep in mind that in the ML world it’s all about garbage in, garbage out and the quality of your data dictates the quality of your outcomes.
Make the most of what you have
Although data is crucial, it is not enough. You also need both hardware and software to process it and get results. It could be tempting to invest in state-of-the-art technology and develop custom solutions, but that is costly and inefficient.
In this game, specialists advise making full use of what is already available. In the software department, try to identify ready-built solutions like libraries, pre-trained neural networks and other modules which you put together like a LEGO set. Sometimes all you need is just an API to connect your back-end to an existing computational solution.
This approach is way faster and cheaper than developing everything from scratch. You need to be aware that the hourly costs for employing machine learning specialists can be very prohibitive for such a project, therefore avoid reinventing the wheel. Instead, take advantage of existing solutions, some of which are even open source.
When it comes to hardware, don’t invest in your servers and tools. Develop your solution around a cloud-based model and find reliable providers. The advantage of this approach is that you can scale as much as needed and pay as you go, keeping your budget in check.
Start simple and scale as you go
The biggest mistake when it comes to machine learning projects is wanting to do everything at once and having the perfect product from the beginning. In fact, this a highly iterative process. You only need to develop a minimal viable product and refine it as you go.
In fact, a good project starts with a simple model which is tested to prove the concept actually works. Next, this model is integrated into business flows and potential misfits managed.
You can only have as many features as you have streams of data available. The overall complexity is also dictated by the amount of recording per each column.
During your development phase, it’s natural to try a few more ideas than you will need or use in the end. Although this discovery mode is essential for the development of the ML model, once you enter a more mature and stable phase it’s important to trim and clean the result, keeping only what’s important.
Final thoughts
Machine learning is a powerful automation tool which can bring projects to a new level, as long as it is used to solve the problems it was designed for. To ensure the success of such an attempt we have listed the top five criteria to take into consideration. The key takeaways include, in the initial phase, clearly defining the project, keeping the profit in mind and identifying data sources. Once the project starts, the focus should be on using all available resources to the maximum and gradually building the complete model. Don’t be afraid to start small.
Guest post courtesy of Maria Weinberger
|