Job Description: Artificial Intelligence Researcher (Final)

Artificial Intelligence Researcher

 

By now, you must have heard of the term artificial intelligence one way or another. For those with a refined taste, you may recall classic films such as 2001: A Space Odyssey or The Terminator. In any case, the moral of the story is the same: don’t trust AI because it is inherently evil towards humans. This certainly makes for a good film plot, but isn’t very realistic. contrary to what mainstream media portrays, a look into my daily responsibilities as an artificial intelligence researcher might help paint a more accurate picture of what AI looks like today.  

 

Artificial intelligence isn’t a specific subject, but a term to describe a broader field of studies. My job is a discipline within AI called natural language processing (NLP). I specialize in the study of relationships between human and computer languages. Ground-breaking as it may sound, I am sure that you are using it every day. For example, mundane tasks like speaking to Siri or searching something on google require some form of NLP.

 

In a nutshell, the responsibilities of an NLP engineer are the following:

·      Prepare

·      Train

·      Improve

Let me use these steps to help explain the process of creating a functional AI agent (another word for an autonomous entity) and what an average day’s work might look like.  

 

Prepare

The preparation process is similar to what you may see chefs doing before service. As an AI researcher, we also go through the process of preparing the data before feeding it to our machines. The steps to do this are:

·      Gather relevant data to train our computer with

·      Clean the data to make sure there are no outliers

·      Convert the data from human language into numbers for the computer

As the name suggests, the preparation process is extremely important because it prepares us for the next step. An error at this stage may not seem significant, but in fact is very costly as we move on.  

 

Train

This is the most interesting part of the process because I can watch the computer improve its accuracy step by step. It’s like watching a kid trying to improve his/her free throw percentage by taking shots over and over again. As I mentioned earlier, this stage can cause major delays if there are errors in the data. These delays happen because the training duration can be very long, sometimes up to several months. Here’s some ways the data might disrupt the training outcome:

·      A corrupt dataset will cause the computer to crash and reset the training process

·      Irrelevant data will waste the computer’s resources and our time. Let’s say we want our computer to translate Spanish, but provide the training data in Portuguese. The computer would be accurate in Portuguese, but that isn’t what we’re after

Having said that, I work tirelessly in the preparation stage to make sure these mistakes don’t happen. At the same time, I don’t usually relax until training is complete. Training may take several weeks and the anxiety might build, but that only adds to the anticipation of seeing the final product.

 

Improve

The last thing that I do is find ways to improve the AI’s accuracy even more. This improvement won’t be drastic, but I can still squeeze a few percentage points out of the system. The things I do at this stage consists of the following:

·      Alter and train the current version of the AI with a smaller and more refined dataset

·      Fine tuning small parts of the algorithms in my code

·      Test the AI again to see if anything has changed and if the AI has reached its limit

While the improvements here aren’t as obvious, the process of fine tuning is necessary. Many times, the two or three percent increase in accuracy may be the difference between a client choosing our AI agent over another company’s. For instance, a client once approached us to build a system that could sort companies into their industries using news headlines. After doing preliminary tests, we told the client we were confident in achieving an accuracy of 94%. I later learned that the client chose us because the 94% we offered was a 4% improvement on their existing system.

 

 

READABILITY REPORT

Flesh Reading Ease: 57.1

Flesh-Kincaid Grade Level: 9.5

Passive Voice: 0%