See the original version of this article on LinkedIn Pulse.
What is Deep Learning?
Deep learning is a subset of machine learning. This machine learning methodology is modelled around the human brain. Our neurons connect to each other through electrical and chemical activations, filtering out different inputs through our dendrites, soma and axons. I know, it sounds complicated. Well, it is. I promise I will try to keep it as simple as possible.
In an attempt to allow a machine to learn like we do, it needs information, very much like we do. Let’s give it the file deep_learning_cat.jpg. It uses a set of “neurons” in its artificial neural network to analyze the image. Each neuron analyzes a different aspect of the file.
The machine will check if the image reaches a favorable result. If not, it will shift the importance it places on the results of a specific neuron over another. This is often referred to as weight. By putting more or less weight on each neuron, it understands better which aspects will be more important to analyze when looking for this deep_learning_cat.jpg.
These weights are generated randomly, and much like evolution, the deep learning machine has to go through many mutations to reach it’s “optimal” point. This machine can then get really good at identifying cats. These neural networks need to have several layers in order to accomplish anything meaningful. These layers are what makes this form of machine learning “deep.” Confused yet? Good. Just read the next section and it will all make sense.
How Humans ‘Deep Learn’?
Answering “What is Deep Learning?” becomes easier when using the most familiar examples, humans. We have been deep learning since we arrived in this world, without ever realizing how cool it is. Attempting to explain how the brain works is a lot harder because to most of us it really ‘just works.’ Muscle memory and practicing instruments are where it gets to be more relatable. We can all easily relate to how complicated it was riding a bicycle with only two wheels the first time.
Now, imagine you play the guitar. You press down on the strings, attempting to form a G chord, only to strum it out of tune and have it go flat half way through. With practice, you will get better as your muscle memory will start to retain certain information. How important is the amount of pressure I place on the strings? The position of my arms? The speed of my strum? The curvature of my fingers?
As you play more, this practice will create a subconscious memory of what is happening, and you place weights on what matters more, focusing on certain things over others. For example, there would be a lot of importance on your eyes looking at the frets of the guitar when you start, but with more experience that becomes less important since your hand can go into perfect position without looking. One human may need 50 practice rounds while the other may need 500, but one can argue that with enough practice any healthy human can produce a perfect G chord.
Now, equate this human importance to the machine’s weighting. The machine would shift its weighting just like we shift our importance, having a high weight on the factors that are more likely to accomplish the task and a low weight for the factors that don’t tend to help the machine accomplish the task.
Let’s look at identifying cats again. One feature that might be important in looking for cats is the eyes, meaning the eye identifier would have a high weight. A less important feature might be the color of the fur, meaning a low weight for the fur identifier. The machine would reach this conclusion after many random selections of weights, altering weights slightly less and less the more likely their current weighting is positively contributing to a specific goal.
When you went through your impromptu guitar lesson moments ago, you practiced playing the G chord a thousand times before retaining some muscle memory. A machine needs practice too, so we feed it big chunks of data before it goes from doing something right 0.01% of the time to doing it right 99.99% of the time – not 100% because there is always room for error.
Where do machines find their data to practice?
Machines find the data needed to practice through the developer’s creation of a data inventory. In most cases, this task is very hands on and depends on the data you need. The more innovative technology gets, the more likely it is the data doesn’t exist and needs to be collected.
With the creation of the desired data inventory, the deep learning machine can start running. As we watch it learn more and more, we anticipate the day that machines learn as fast as – or faster than – us.
Share any questions and give us your insights about the topic of “What is Deep Learning?” below.