3 minutes reading time (679 words)

AI & Accuracy: How to Keep Systems Honest

Earlier this year, Stanford and Google found that a machine learning agent designed to transform aerial images into street maps and back was cheating. Turns out it was hiding information it would later need in an almost undetectable high-frequency signal that developers didn't immediately catch, and was only suspected because the results seemed a little too perfect.

While a fun anecdote about Artificial Intelligence (AI) development, it still raises the question of how developers can ensure their systems are properly executing the tasks they're meant to. Or rather, as was the case with Stanford and Google's AI, how can developers establish tools and criteria to thoroughly evaluate AI and machine learning systems?

Before jumping into this, there are two key elements to consider:

Inaccurately Interpreting Research Reports and Wrongly Assigning Intent

 

As exciting as the prospects of AI are, it's important not to overestimate the actual state of possibilities. A careful review of the research projects may reveal that some of the more amazing results actually leave out a lot of circumstantial data, and that replication is difficult to achieve at this point in time.

Likewise, media plays a part in inflating our expectations of AI technology, but it's not necessarily rooted in capabilities that are available to us today. No matter how advanced a system may appear, it's important to remember that computers don't possess their own will or motives, and can only execute the tasks we ask of them—even though they can learn, recognize patterns and make adjustments based on the information they receive.

Leveraging Machine Learning and AI Systems with Accuracy in Mind

 

 So, how can developers make sure their results are accurate? Here are a few guidelines to follow:

  • Understand AI limitations and adjust expectations
    Keep in mind that AI initiatives are technology projects, not magical genies. Their performance will depend on a business context to solve problems and produce results. This means developers need to set and focus on a business goal to properly direct efforts and identify how the AI can best help achieve their desired outcomes.

  • Clearly define roles for both human and AI sides
    We're not at a point where AI is ready to think for us. That's why a human touch is still vital, especially when it comes to decision-making. AI can easily run through large amounts of data and make recommendations, but human engineers should always be the ones helming the ship.

  • Evaluate existing processes
    Run AI model simulations continuously to confirm your AI's repeatability. If you aren't coming up with consistent results, your AI might not be ready for launch. At the same time, examine your processes with a critical eye to catch any blind spots.

  • Check data quality
    Your system's analysis will depend on the quality of your data. Poor data means poor results, so remain watchful of the kind of data that's being analyzed.

  • Understand the model's process
    You should be able to trace the model's "thought process" and explain how it reached any given conclusion. This makes it easier to understand where any mistakes or errors are coming up, and how to solve them moving forward.

  • Continually add new data and retrain your AI

An outdated system is never going to yield good results. Keep your system up to speed by including the latest data, more examples, more features, etc.

  • Keep an override mechanism handy, just in case
    It's a good idea to have a manual override mechanism available in case of software, hardware, or network failure, even for non-critical operations.

    Clearly, there's still a long road ahead before sci-fi level artificial intelligence systems are ready to take over human jobs or roles. But, to the extent that developers maintain rigorous standards and keep evaluating with clear goals in mind, systems can become better and more accurate at performing their tasks.


--
How can your AI and machine learning projects benefit from careful attention to detail and a commitment to quality? Find yourself a great partner! Let's schedule a chat to find out if PSL is the right nearshore outsourcing partner for you.

PSL's Daily Standup with Juan Ruiz
Using Machine Learning to Detect Anomalies in Perf...

Related Posts

 

Comments

No comments made yet. Be the first to submit a comment
Already Registered? Login Here
Guest
Friday, 19 July 2019

Captcha Image