If you follow the jargon in AI, we have a new entrant- Liquid Neural Networks (LNNs). And if you dig deeper into this new avatar of neural networks, you will find out why this avatar can be very beneficial in the business world. Specifically in the context of autonomous planning, autonomous supply chains, or (almost) self-running enterprises.

To stay out of my comfort zone, I will use an example of marketing to illustrate how LNNs can better fit in the context of autonomous business processes.

The “liquidity” approach of neural networks can be extrapolated to many other AI algorithms, not just neural networks. But let us try to understand this liquidity from the perspective of NNs.

You have developed a deep learning algorithm for hyper-personalized targeting of customers as they walk through your retail stores. With their consent, you have used their buying behavior, demographics, online browsing history, store floor browsing routes, real-time store movement, and our current store layout and assortment to develop a hyper-personalization NN. 

This hyper-personalization NN sends a text to a customer who has just walked into the cleaning products aisle and has shopped for dishwashing pods during that week of the month, for every week, that an offer was available on their favorite brand if they buy a bigger pack.

Surprisingly, the customer ended up not buying the product this month. And if the NN accounted for their most recent browsing data that the customer consented to share, they would have found they browsed for a different brand on their website. From there, they switched to a competing e-commerce portal. Probably to compare prices.

They probably found a better deal somewhere else. Hence, they decided not to buy this product at your store. Then, for the next three months, they repeat the same pattern.

Your NN now needs to account for this behavior. But it could not do that for the next few months because it was not “upgraded” with the most recent data. 

This is where the training aspect of NNs comes into play.

We currently deploy these models by training the models and then deploying the most recent version. The parallel models keep training on new data. The deployed model gets updated with more recent versions from time to time.

Liquid NNs aim to eliminate this decoupling. By making a NN train as well while it is in production.

And the real value lies in eliminating the term “time-to-time” in one of the paragraphs above.

LNNs will learn “on the job”. So the latency, illustrated in the previous example, will not lead to missed opportunities like the example. But the fact is, many of our visions of autonomous enterprises, which can certainly be brought to life by NNs, can only become a reality if we can empower NNs to both learn and make decisions with minimal latency.

This looks promising. And again, it seems to keep the exponential progress of technology and AI algorithms on track.