in Continuous Feedback



As humans, we learn in a more fluid and robust manner than computers have been able to, thus far from many basic recognition tasks. Many problems that challenge computers even today are trivial for humans. We easily recognize faces in varying lighting conditions, orientations, and more, even when seeing with them for the first time. And we generally use the available data intelligently—ignoring superfluous or unnecessary data when it is plentiful and making the best of it when it is in short supply. Being impervious to rotation and scale which organisms as primitive as ants have employed with great success in learning about danger or finding useful supplies.


Deep learning is a lot of things, but one thing’s for sure- it isn’t easy and as fancy that it looks, even data scientists who have mastered the basics of artificial neural networks and intelligence, they might need time to level up and be on the same page on the intricacies of the multilayer deep learning algorithm.

Still confused what deep learning exactly is?

Well, let’s put it in simple terms. Deep Learning is a machine learning technique which helps the computers to do what comes naturally to the humans. A computer model that learns how to perform classification tasks directly from images, text, or sound effortlessly. Deep learning models have the ability to even exceed human-level performance sometimes. It attains such impressive results with one golden quality embedded in them— accuracy. Deep learning has been able to achieve recognition accuracy at higher levels than ever before and the below diagram depicts the errors being made by deep learning pre and post 2012.

One of the examples of deep learning could be in the hearing and listening translation aspect. Those home assistants which respond to your voice and know your preferences are deeply rooted within the deep learning applications. Well, a classic everyday example could be your phone’s Face ID. When setting up your phone, your face is scanned and thousand data points are collected by the algorithm. A thorough depth map is created which thereby helps the engine to decide whether it is you or not.

The recent AI technology enabled self driving cars are also backed up with the concept of deep learning. Automotive researchers are initially responsible for classifying the objects into categories like the traffic signal, sign board, detect pedestrians, wrong ways and tactics to decrease accidents. It is very important to meet the expectations of the users for such safety-critical applications.


Reinforcement Learning is primarily based on the concept of reward hypothesis. It refers to a goal oriented algorithm which learns not only how to attain it but also how to maximize it over many steps. These steps are nothing but essentially through trial and error methods.  Therefore, in simple terms, it is learning by doing in order to achieve the best outcomes. 

In the real world, one of the easiest and most common examples to relate Reinforcement Learning is by understanding the learning process of a baby. When a child in the living room sees a fireplace and approaches it. He feels the warmth of the fireplace, it’s positive for him and thereby feels good (Positive Reward +1).  He now understands that fire is a positive thing. But then when he tries to touch the fire, ouch! It burns his hand (Negative reward -1). He’s just understood that fire is positive when you’re at a specific distance away, because it produces warmth. But if he gets too close to the fireplace, it may cause him harm.

Well, that’s how humans learn in the end, through interaction. Reinforcement Learning is merely a computational approach of learning from action to be taught. I believe games can explain the concept of Reinforcement Learning problem in the best possible way. Let’s take the game of PacMan where the goal of the PacMan is to eat the food through the grids and at the same time to avoid the ghosts on its way. PacMan receives a reward for eating food and also receives a punishment if it gets killed by the ghost. It ends up losing the game. The total cumulative reward is PacMan winning the game. Simple, isn’t it? 

Being one of the hottest research topics, Reinforcement Learning has been proved to achieve the kind of performance handling complicated tasks smoothly and efficiently. Horizon, a reinforcement learning platform by Facebook, is addressing and overcoming the challenges in this area. It’s an open source and an end-to-end platform that facilitates Reinforcement Learning. Facebook the ever dominating social network platform has yet again proved that it can contribute to the enhancements in Machine Learning algorithms. Facebook’s Horizon approach regarding

Reinforcement learning algorithms can be expected to perform better and probably the best in more ambiguous and real-life environments rather than from the limited options of a video game. That is, with time we expect them to be valuable to achieve goals in the real world.

Both of them, deep learning and reinforcement learning are ]systems that learn autonomously. The major difference between them is that deep learning is learning from set 1 and then applying that learning to set 2, while reinforcement learning is dynamically learning in the same environment by adjusting actions, based on the continuous feedback it gains through repeated attempts.


Write a Comment