Enterprise-level companies, with thousands of employees across the globe, can have their AI framework replicated by B2B companies of nearly any size. In fact, even without access to a ton of proprietary data and immense computational power, you can build out effective AI tools.
To begin with, you can introduce publicly available datasets from movie review websites, e-commerce review sites and elsewhere to train your system. When doing so, be sure to double check the license details of such publicly available corpuses, as some will not allow commercial usage. Then, during the preprocessing stage—after removing swear words, jargon and other irrelevant elements from your corpus—you can add noise and slightly alter your existing data to create even more relevant data points in your corpus.
After this, you can build a semi-supervised learning algorithm through techniques such as tri-training. In tri-training, we essentially train three different neural networks over binary classification data. Once two models agree with one another, the data is introduced to the third network and then re-trained until all three models are in agreement. However, before your AI tools are up and running, it’s important to be mindful of the following common pitfalls.
Overfitting AI Models
It may seem counterintuitive, but you really do not want your AI model to be too accurate. If your cross validation score is anything above 100, then you might need to reconsider deploying the model, since it might be prone to overfitting. If your training data is overfit, one small change can render the entire model useless. Your model should be accurate. However, it cannot be too accurate, as it needs to be adaptive.
Class Imbalance
When training your AI tools, it sometimes makes sense to include a disproportionate amount of your target variable. For example, perhaps you’re training your computer vision AI tool to recognize planks of wood in a warehouse. Even if planks of wood only account for 1% of your total inventory, you’ll want to oversample the minority class. In this case, you’ll likely want 60% of your sample to be planks of woods. You really want to amplify this number to ensure you have the right class balance.
Data Leakage
Another common misstep is the failure to capture data leakage, which can occur in any number of ways. A common instance occurs when an element used to test the model slips into the training data. This generally causes your model to appear to be performing well. However, its predictions might not be as good when deployed in production. To rectify this, it is vital that you keep track of any changes to your AI tools. In addition to your organization’s code reviews and configuration change reviews, you should have data set reviews to ensure your AI engineering is being done properly. Any parameter change within your AI tools needs to be fastidiously documented.
Concept Drift
All AI tools are dependent upon the shape of the data that comes into your system. For example, perhaps your AI model was trained for response times being between 1-100 seconds, and all of a sudden the response times are between 1-1,000 seconds. In this case, your model will likely fail to capture the new data points. It’s important to be aware of such an occurrence, so that you can retrain your model accordingly.
There are many techniques that can be used to identify instances of concept drift. Rather than waiting for feedback from your customers, you can set up anomaly detection in a way that it identifies concept drift on incoming data and automatically updates your AI model.
Using Personal Data
With the CCPA and GDPR in place and a federally mandated U.S. data privacy law on the horizon, organizations are wary of using their customers’ data to train their AI tools. It’s prudent to train your AI system with anonymized data; this way, you won’t set yourself up for problems down the road.
Failing to Ensure All Your AI Decisions Are Explainable
Most AI-enabled decisions are enacted upon by teams rather than individuals. Hence, it’s vital that your AI model’s recommended courses of action come with explanations. Additionally, these explanations should have confidence intervals, making it easier to interpret results.
During root cause analysis and remediation, it’s often unclear which team is ultimately responsible for the issue at hand. If the AI’s suggested course of action comes with an explanation, this helps immensely during the remediation process. For example, if the starter event is a network build-up, the issue likely lies with the engineering team. Alternatively, if the starter event is a network configuration change, the network team is the likely culprit. By closely monitoring your AI tools’ performance and ensuring that all AI decisions are explainable, you can quickly solve issues and course correct.
Guy Fighel, general manager and group vice president for New Relic, said these and other capabilities being added to extend the capabilities of a monitoring platform that as a software-as-a-service (SaaS) platform provides a natural focal point applying machine learning algorithms to the massive amounts of data required to inform an AI engine.
Organizations that are finding it difficult to handle the volume of support calls being made by often geographically distributed remote workers can employ chatbots via a SaaS platform in a few hours, Calhoun said. In many cases, IT staff are trying to respond to help desk requests using on-premises applications that they need to access remotely, often using cumbersome virtual private network (VPN) software. A SaaS platform provides an alternative approach that is designed from the ground up to be accessed remotely, he said.
