Deep learning amazed the audience when Google’s AI beat a human top player at the game of Go, an ancient contest, and a game of strategy that oppressed many AI experts from the past decades.
The robot player became an overnight sensation.
Machines have been well known to top the best humans at games that involved human intellect – Scrabble, Chess, Othello, and Jeopardy. Well, we’re not comparing Go to be like any other games. Go has been a 2500-year-old known for its unmatched complexity.
All thanks to AI’s technology – deep learning.
Deep learning: what is it?
In practical terms, we can define deep learning to be a subset of machine learning. Technically speaking, deep learning is machine learning and can function the same way. However, both hold different capabilities. Let’s not confuse these terms.
Machine learning models get better with time; however, they still require some sort of guidance. For instance, if an AI algorithm runs an inaccurate prediction, an AI engineer needs to give in a hand, adjust the model to properly function again. Whereas, with deep learning, the algorithm can determine the prediction stating whether it is accurate or not with the help of its neural network. For example, with the help of multi-layered artificial neural networks, accuracy in tasks such as speech recognition, identifying an object, and language translation can be easily performed.
Difference between deep learning, machine learning, and AI
Artificial intelligence, deep learning, and machine learning are terms that often overlap.
Here’s a simple way to make you understand.
- Artificial Intelligence simply means getting your computer mimic human behavior in some way or the other.
- Machine learning is a subset of AI that involves techniques enabling the computer to learn from the data and further deliver its application.
- Deep learning is a subset of machine learning that helps the computer solve much more complex problems.
Understanding the latest inventions in AI can get overwhelming, but if you’re clear with the simple logic – you will easily understand the differences and their applications.
According to Ray Kurtzweil, an American inventor and a futurist predicted stating AI singularity to happen by 2045.
As he defines, “the moment when a 1000$ computer may contain as much computing power as 1000x the human brain has.”
Ray is confident this will soon take place. This is where he insisted he would be needing to work harder to reach the true singularity i.e. better algorithms.
In short, perhaps we still lack finding the best mathematical formulas. Therefore, till then for learning to properly take place using deep learning, one needs to first feed a large amount of data to the deep learning algorithms.
The Why’s and the How’s
Why deep learning matters?
One word: accuracy.
Deep learning achieves accuracy at a higher level. This in return meets the user expectations, it also plays a major role in safety applications in driverless cars. Recent advancements in deep learning have proven to outperform humans at tasks involving classifying objects from images.
Though deep learning was first defined in the 80s, it is only till today it has become useful. How?
- It required large amounts of labeled data. For instance, to develop a driverless car, it would require millions of images along with thousands of hours of video.
- A substantial amount of computing power is needed. Thus said, high powered GPUs tend to have a parallel architecture that is crucial for deep learning. in combination with cloud computing, it can easily reduce the training time for the deep learning network to function – the reduced time from weeks to hours or maybe even less than hours.
How does it work?
Most of the deep learning method uses a neural network architecture, the major reason why deep learning is referred to deep neural networks.
“Deep” means the number of hidden layers in the neural network. Even so, a deep network may contain more than 150 layers whereas a traditional network can have only 2 to 3 hidden layers.
Now, these neural networks composed of layers of nodes, in the same manner as the human brain consisting of neurons. The nodes within the individual layer are connected to adjacent layers. The network is considered deeper based on the number of layers they possess. In a human brain, a single neuron can receive thousands of signals from other neurons. Following the same pattern, signals that have traveled between nodes and have assigned corresponding weights in an artificial neural network, it is said that the heavier the weight of the node is the more effective it will have on the next layer of the node.
So, the last layer compiles all the weighted inputs to produce an output. Since deep learning uses large volumes of data that need further processing that involves severe mathematical calculations, they would require powerful hardware.
When huge data sets are fed in the deep learning system, with the help of artificial neural network these data can be classified with the answers received from the series of binary true or false questions. For instance, a facial recognition functions when it has learned to detect and recognize the lines in the face, then the more significant part followed by the overall representation of the face.
Applications of deep learning
AI is getting smart day by day and why not. With the amount of computational power, it possesses, machines have the capability of recognizing objects and translating them in real-time. Let us look at the top deep learning applications widely used today.
- Voice search and voice-activated assistants
- Self-driving or driverless cars
- Automatic machine translation
- Automatically adding sound to silent movies
- Automatic handwriting generation
- Automatic text generation
- Image recognition
- Predicting earthquakes
- Automatic colorization
- Neural networks in finance
- Neural networks for detecting brain cancer
- Automatic image caption generation
Deep learning is all about intelligence.
In the foreseeable future, deep learning will be more about spiking neural networks.
