[av_textblock size=” font_color=” color=”]
Our CTO, Dan Raviv, recently wrote a guest column for Forbes highlighting ML Algorithm. ( https://www.forbes.com/sites/forbestechcouncil/2018/11/26/how-can-you-build-a-winning-ml-algorithm/#630a870d2da0)
The impressive and rapid development of machine learning (ML) algorithms, especially those that are based on deep neural networks, has made artificial intelligence (AI) a mandatory technology in the 21st century. Such tools are advantageous across a wide range of industries, including — but not limited to — e-commerce, education, retail and finance, and are thus relevant to many companies.
In essence, neural networks consist of many neuron layers stacked together in some configuration. The connection weights between these layers are trained by the backpropagation algorithm while minimizing a specific cost function. This framework happens to provide state-of-the-art results; however, the underlying mechanism of such data-driven solutions still puzzles researchers and is an active area of research.
How can you build a winning ML algorithm for your company? There are three main options:
Go The Open-Source Route
The first one — and the cheapest — is to use open-source free packages for learning. Run some experiments with your data, and choose the method that performs the best. In many cases, this is good enough and can even provide adequate results. Not surprisingly, many tasks share a lot in common, and pre-built models, and even pre-trained networks, work satisfactorily out of the box. Sometimes, however, one needs to retrain a part of the network, add one or two new layers or adjust some outputs — a procedure better known as transfer learning — but the desired result is achieved with minimal effort.
Several companies offer AI services or algorithmic solutions based on ML. This means they inject their code into yours, open an API for learning and real-time predictions or provide SaaS services. Such companies have a team of AI experts who fine-tune their tools for your needs. Usually, it is done once for the data you provide, where future optimization or further development increase the cost.
Have A Team Build A Customized Algorithm
The last option is to hire a team of experts who can build an in-house AI platform designed specifically just for your company. This option is by far the most expensive, but it’s also the one that typically can provide the best results due to the customization factor. Getting an AI system up and running is not a one-and-done procedure. It is a dynamic process that needs care and attention. Features are constantly changing, cost functions are updating, requirements from the business side are shifting, and new methods and algorithms are published on a daily basis. More importantly, the problem you wish to address with AI may be unique such that none of the existing solutions would be useful.
What About AutoML?
In an effort to try to close the gap between a dedicated team and an out-of-the-box solution, a new field has emerged known as automated machine learning or AutoML. This new approach tackles the problem of choosing the right AI architecture for a specific problem. The backpropagation algorithm, which is based on the derivation of a cost function, is used to optimize the connecting weights, but neural networks have a lot of other knobs to turn. Although updating the topology or intelligently optimizing some hyper-parameters showed improvements in many models, their success is still limited. AutoML methods start with an initial guess, or even with a known state-of-the-art solution that has proven useful for other tasks, and aim to systematically explore numerous models in the search for the best one given the specific goal.
Should data scientists fear for their jobs? The answer is an emphatic no. There will always be a gap between the results provided by well-trained data scientists and the results provided by an automatic or semi-automatic optimization procedures. Every problem is unique and the ability to foresee all is far beyond the scope of such systems. The great minds that are invested in the next generation AutoML systems would also be able to push the performance even further given a specific task. Having said that, such automatic tools will help experts to better fine-tune AI solutions and squeeze out more out of existing ones. Not only will that save time and money, but it also would allow experts to devote more of their time for original and exciting methods for the problems we face ahead.
Are you new to the US? Need a car loan? Lendbuzz can help you purchase the car you need. We offer competitive rates for international students, professionals, and expats. No credit score, social security number or co-signer needed. We help you through every step of the way. From applying from your home country to picking the right car for you. Apply online in minutes and get your rates today.