While the coronavirus pandemic has rearranged our plans and pushed us back into our homes for extended periods of time, people have found new ways to stay entertained and connected remotely. Although users might not see the artificial intelligence applications used in their online entertainment, shopping and social networking platforms, AI informs much of what we see, interact with and purchase every day.
AI might seem like a futuristic concept, but the truth is, 85% of Americans are already using products powered by the technology. Social media platforms such as TikTok have catapulted into popularity, while many home-streaming networks saw an increase in subscriptions. Similarly, the social-networking platform LinkedIn recorded a 26% increase in user sessions between January 2020 and April 2020.
To streamline their expansive databases, platforms like these can record your preferences and use AI to show you the information, videos and products it thinks you’d like best. Ultimately, AI is leveraged to transform digital experiences, and with investments growing around the world — to the tune of $25.2 billion in investments in American companies alone — I expect this trend to increase, even after widespread vaccination.
As the CEO of a company that combines human expertise with machines to generate accurate data for AI models, I’m always paying attention to AI and its role in our lives. I believe that as companies continue to use deep learning algorithms to power our favorite platforms, our collective responsibility as leaders in tech is to ensure the applications are built using high-quality, unbiased data. Without accurate inputs, AI applications won’t be effective and could even produce potentially dangerous outcomes.
Here are prominent examples of how AI is working in the background of our everyday lives:
• Personalized shopping experiences: Some online marketplaces use an internal recommendation system to help boost on-site sales. Amazon, for example, uses machine learning to match shoppers “with other customers who had similar purchase histories, and those purchase histories would suggest recommendations for the visitor.”
• Social media: Social media platforms use AI to surface content they believe viewers would be most interested in based on their preferences and historical clicks. Creators and small and midsize businesses rely on AI technologies within these platforms to connect with their followers. Content automation tools also support creators with suggested design templates, creating content that aligns with their brand aesthetic.
• Job searches: LinkedIn, similar to other social media platforms, uses AI to surface content most relevant to its users, suggest connections and serve relevant ads. It also uses AI to group similar positions that might have different titles, identifies the skills that the job requires, then suggests the job to a LinkedIn member with relevant experience.
The key to ambitious AI applications is high-quality and accurate data.
To guarantee these AI innovations are effective, leaders in tech must ensure that the data powering the technologies is high-quality. With unbiased algorithms and clean data sets, the possibilities for AI are endless.
So, what’s the key to unlocking successful AI? There are a few things for leaders to consider.
First on the list is leveraging machine-learning-assisted annotation effectively. My company uses ML-assisted annotation tools, and through this experience, I’ve found that this tech can enable companies to create a scalable and high-quality feedback loop that begins with humans training the machine, followed by machine assistance, and is then repeated. These applications can also help reduce the dangers of algorithm biases.
To ensure efficacy, business leaders should consider some best practices when incorporating ML-assisted annotation. I’ve found it most beneficial to prioritize growing and nurturing ML product teams. Humans checking for accuracy is a crucial component in the quality and performance of AI algorithms. With that in mind, product leaders and managers should consistently have training sessions and all-hand team meetings to encourage continued learning and skill improvement.
Beyond using ML-assisted annotation, leaders can also use specific strategies to counter bias in training data. One of those is to avoid sample bias and ensure reality is represented. It’s vital to source and select training data with the end result in mind to avoid sample bias and class imbalance. If you want to identify pedestrians, for example, source city street data showing people from all demographics. Highway data with few people won’t help you ensure unbiased data in your algorithms. And refresh your data several times a year as the world changes.
Also, be sure to test your algorithm for bias before training with dataset tests. This includes doing things like cross dataset generalization, where you test a model trained on one data set to another set. This works best for object detection models. Additionally, you should conduct a negative bias test, in which you test a set with positive classifiers from its own set but negatives from many data sets. This shows if positives are working on specific objects and not the whole scene.
Another solution is to divide the problem, or your algorithm, into sub-problems. To do this, you can train a few localized classes in rounds, instead of all objects in the training data at once. For example, if you’re training an algorithm to recognize images in a retail online site, you can train all shoe types first, then move onto pants, etc., as opposed to training the entire catalog at once.
So the next time you’re scrolling through shopping on Instagram or keeping an eye out for your next opportunity on LinkedIn, put your platforms to the test. See how accurate your personalization is or isn’t because it all comes down to the quality and accuracy of the data that’s feeding the platforms’ AI models.