AI bots are developed using historical data sets. These bots are good at prediction, but they can’t cope if your data constantly changes. That’s when you need to adopt continuous learning.
The Coronavirus crisis has had a big impact on everyone’s lives. In particular, it has a big impact economically. Small businesses are shutting down, millions of people are being laid off, and governments are facing a fiscal black hole. Huge events like this demonstrate the weakness in most economic models. Even the best models built with the latest AI suffer because they can only use historical data for their forecasts. But one AI approach offers a solution—continuous learning. Here, I explain what that is and how we leverage it at Sonasoft.
What is continuous learning?
As humans, we are continually learning from our surroundings. Spotting patterns in events, gaining new experiences, and learning from our mistakes. The human brain is incredibly good at making connections that are the key to learning. As a child, you quickly learn that if you touch something hot, it hurts. The next time, you are a bit more cautious and check how hot it is before you touch it.
Children often get things wrong though. For instance, your daughter might like oranges. But she’ll probably get a shock when she mistakes a grapefruit segment for a piece of orange. Her brain will now store away this new information that oranges and grapefruits taste very different.
This continual process of learning by experience is one of the main aspects of human intelligence. The other is our ability to make leaps of imagination and to create new ideas.
How does AI usually work?
Artificial intelligence is sometimes a bit of a misnomer. All too often, AI is used to describe systems that incorporate a simple machine learning model. These ML models are trained on completely static datasets. The data scientists have taught the model to spot certain patterns in the data. Usually, these same data scientists have engineered the data to highlight the features that they are most interested in.
Models like these are examples of narrow artificial intelligence. They are powerful, and can often perform their task better than us humans. However, their static nature means they are thrown as soon as anything fundamental changes. Let’s look at a simple example.
Imagine you have created a model looking at the demand for Internet access across a city. To do this, you have collected records of internet usage over the past year and trained an AI model. Your model probably identifies a clear diurnal pattern. Daytime Monday-Friday, the majority of the demand comes from downtown. In the evenings, the demand comes from residential districts, peaking around 8-10 pm. At weekends, demand drops a bit as people head to their weekend homes. But now imagine how inaccurate your model has become since the pandemic started. All of a sudden, downtown has got really quiet, and home usage has soared. This is when you need to adopt a continuous learning approach.
What is continuous learning?
In continuous learning, your machine learning model is dynamic. In a normal model, new data is used to create a prediction. However, the model itself remains static. This means that it can’t adapt to changing circumstances. But continuous learning models are different. The model itself is updated every time you add more data. Importantly, this process happens smoothly and without interruption.
Continuous learning is closely related to reinforcement learning. However, they aren’t completely synonymous. Reinforcement learning is certainly a form of continuous learning, but it isn’t the only approach. Reinforcement learning typically involves deep learning models that evolve over time, based on the feedback they receive. But you can also create models that combine the new data and predictions without needing a deep neural network.
There are three forms of continuous learning. The simplest grows the training data but doesn’t add new items. For instance, if your model is distinguishing photos of dogs, you can add more dog photos in different poses and lighting conditions. Alternatively, you might expand the scope of the model. For instance, teaching it to also recognize cats as well as dogs. Finally, you can combine both these, providing the model with photos of new subjects and new versions of existing ones. This last form is the most powerful but is also the hardest.
How Sonasoft leverages continuous learning
Here at Sonasoft, we have developed an AI bot factory called Saibre. NuGene is designed to make it extremely easy to develop AI bots for business-critical applications. There are several things that set Saibre apart from the competition. These include the range of data that can be processed, the understanding of time-series data, and checking for causality. One of the biggest differences is the application of continuous learning.
can create a wide range of different bots to help businesses with forecasting, knowledge discovery, anomaly detection, and more. To start with, you provide as much historical data as you can. Depending on your use case, this could include numerical data, time-series data, video, audio, and even IoT sensor data. Saibre will analyze the data and try to find interesting correlations. From these, it will develop hypotheses, which it then checks for causation. Once the system is sure that its hypotheses are correct, it will start to train a suitable AI model. The model is verified and then embedded into an autonomous AI bot. Up to this point, Saibre is replicating what a competent team of data scientists would do.
The important difference is that Saibre also applies continuous learning. While your AI bot is operating, Saibre is collecting all your new data and checking every prediction or model output. It constantly retrains the model based on the new data, ensuring it stays up to date.
A real-world example
Recently, Sonasoft signed a deal with an East coast electricity cooperative. We are providing an AI bot that accurately forecasts demand. These forecasts allow the cooperative to trigger load control when demand is expected to peak. The model combines multiple factors, including historical demand, future weather predictions, and knowledge of events like Superbowl. From these factors, it predicts both the peak and the best time to implement load control. This results in significant savings since the peak demand determines wholesale electricity prices.
This bot was trained on historical data and then verified on data from 2019. Once the bot goes live, it will use continuous learning to constantly update its models. This will ensure the models are able to respond to changes in demand. As a result, the models will remain accurate despite events like the pandemic that lead to major changes in demand.
What next for continuous learning?
Many academics say that continuous learning is an important step towards developing true artificial intelligence. This is because such models are much closer analog to how a human learns. Whatever the future, there’s no doubt that continuous learning overcomes the well-known issues with ML models becoming stale. Here at Sonasoft, we are constantly updating and improving Saibre. The result is a system that applies continuous learning both to improve its own performance and to improve its bots.
AI bots are typically static. This means they quickly become outdated as your business evolves. The solution is to apply continuous learning.