Understanding Artificial Intelligence Technologies - Deep Learning and Reinforcement Learning

The industry is witnessing too much of technological advances especially in the field of Artificial Intelligence. The disruption has been so overwhelming that there is no dearth of professionals who aren’t awestruck by the awesomeness and rise of the exponential technologies. Enterprises with ever-growing business demands are beginning to dig deeper into the findings and services of technology leaders like Amazon, Google and Microsoft. Therefore, the various emerging technologies belonging to the AI portfolio are gaining a lot of importance and being discussed as well as tried out by stalwarts belonging to different fields.

As the amount of data business houses generate continues to grow to massive mind-boggling levels, the AI maturity and the potential problems AI can help solve grow right along. The huge volumes of data and the amazing computing power that’s now available for a reasonable cost is what triggers the tremendous growth in Artificial Intelligence technologies and makes deep learning as well as reinforcement learning possible.

The velocity with which AI industry is undergoing development, it can be challenging to keep up with the latest cutting-edge technologies. This post is an effort from my end to provide a simple and clear understanding on deep learning and reinforcement learning so that you can figure out the difference. Both deep learning and reinforcement learning are machine learning modules, which in turn are part of a wider set of artificial intelligence suite.

Why are deep learning and reinforcement learning modules captivating our interest and what makes them stand out amongst other AI tools? It is their unique ability to empower a smart device to develop rules on its own to solve problems. This ability to learn is not nascent for smart devices, but until recently we didn’t have the data or computing power to make use in this manner.

Deep Learning is necessarily an autonomous, self-teaching system in which you use existing data to train algorithms to find patterns. You then use these patterns to make predictions about new data. For example, you might train a deep learning algorithm to recognize lions on a photograph. You would do that by feeding it millions of images that either contains lions or not. The program will then establish patterns by classifying and clustering the image data such as edges, colours, shapes, distances between the shapes, etc.). Those patterns will then inform a predictive model that is able to look at a new set of images and predict whether they contain lions or not, based on the model it has created using the training data.

Deep learning algorithms work through various layers of artificial neural networks which mimic the network of neurons in our brain. The neural networks allow the algorithms to perform various cycles to narrow down patterns and improve the predictions with each cycle. A fantastic example of deep learning in practice is Google’s Fingerprint Recognition. When setting up your phone you train the algorithm by scanning your finger(s). Each time you log on using e.g. Fingerprint ID, the fingerprint camera app captures thousands of data points which create a depth map of your fingerprint(s) and the phone’s inbuilt neural engine performs the analysis to predict whether it is you or not.

Reinforcement Learning is an autonomous, self-teaching system that necessarily learns by trial and error. It performs actions with the objective to provide maximum benefits, or in other words, it is learning by doing in order to yield the best results. This is similar to how we learn things like swimming where in the beginning we keep going out of breath, end up drinking water and make too heavy and often peculiar moves with no streamlining technique in place. However, over time we use the feedback of what worked and what didn’t to fine-tune our actions and learn how to stay afloat, finally swimming in proper fashion. The same is true when computers use reinforcement learning, they try different actions, learn from the feedback whether that action delivered a better result, and then reinforce the actions that worked, i.e. reworking and modifying its algorithms autonomously over several iterations until it makes decisions that produce the best result.

A valid example of using reinforcement learning is a robot learning how to walk. The robot first tries a large step forward and falls. The outcome of a fall with that big step is a data point the reinforcement learning system responds to. Since the feedback was negative, a fall, the system adjusts the action to try a smaller step. The robot is able to move forward. This is an example of reinforcement learning in action.

One of the most fascinating examples of reinforcement learning in action was when Google’s Deep Mind applied the tool to classic Atari computer games such as Break Out. The goal (or reward) was to maximize the score and the actions were to move the bar at the bottom of the screen to bounce the playing ball back up to break the bricks at the top of the screen. In the beginning, the algorithm makes lots of mistakes but quickly improves to a stage where it would beat even the best human players.

So how does deep learning differ from reinforcement learning? Deep learning and reinforcement learning are both systems that learn autonomously. The difference between them is that deep learning is learning from a training set and then applying that learning to a new data set, while reinforcement learning is dynamically learning by adjusting actions based on continuous feedback to yield the best result. Deep learning and reinforcement learning aren’t mutually exclusive. In fact, you might use deep learning in a reinforcement learning system, which is referred to as deep reinforcement learning and I will discuss it in another post.

You can connect with me @Asamanyakm