Deep Networks – Drop a Pebble in a Pool

Machine learning and artificial intelligence are advancing quickly!

A lot of professionals recognize this, rightly, as the new frontier of computing. We’ve already come up with sophisticated big data strategies to get business insights. The next step, it seems, is to get computers to “learn” and “think” like us – to do more of the cognitive work on their own, and to bring us even more elaborate knowledge and business data based on their high-powered artificial intelligence platforms.

Echo State Networks and Liquid State Machines

Some types of deep learning networks now have specific builds that professionals can talk about when they discuss exactly how intelligent machines handle large sets of data.

One of these is the echo state machine – and its corollary, the liquid state machine. In this case, you can make an intuitive analogy and explain the work of these neural networks this way.

By dropping a pebble into a pool of still water, you create a ripple effect that is generally predictable but statistically complex.

What if you drop two pebbles at the same time? Then the ripple effect becomes harder to predict and more complex, the same way machine learning programs digest more complex sets of unlabeled data.

The IT community has already learned to talk about data in terms of water – for example, with “data lakes” and “data pools.” However, this water analogy takes things a lot further.

One way to describe this is that the neural networks involved in these types of machine learning programs have sets of weighted inputs that lead to an activation function, or more accurately, a set of activation functions. Information travels through the layers of the network and comes out as output. Think of the act of throwing the pebble as the act that generates the inputs, and the evaluation of ripples as the outputs. Through backpropagation strategies, the network can go back and figure out how those two things relate.

Transparency and Opacity

This kind of chaining back through the system to link the inputs to the outputs can be done to a certain extent – but there’s also a kind of black box model in play. By building neural networks in more and more complicated ways, scientists lose some of the ability to really look at how they work directly.

Some types of unsupervised machine learning are really not really open to explanation. Engineers can build them and make them work – but they can’t really explain everything about how the machines come to their conclusions.

The bottom line is that all of this is very exciting on a technological level. It also has direct applications to business. Recommendation engines, customized marketing and in-store virtual-reality are just the start. We’re entering an age when companies can learn more about their customers, go deeper in delivering products and services, and really cash in on smart computing. You’re going to see the results fairly soon.

Let WebSubstance help guide you to utilize big data and AI to stay ahead in competition in today’s fast-moving business world.