What is AI Machine Learning?
AI machine learning is important to understand when considering how computers function. AI is an acronym for artificial intelligence. It is an umbrella term that covers the science of teaching computers to think. This is a very simplistic way of defining AI machine learning.
So, the science of teaching computers to make decisions is artificial intelligence. Machine learning is one way that this is being done.
Now, there is also reinforcement learning. Reinforcement learning is another branch of artificial intelligence. Machine learning is now broken down into supervised, unsupervised, or semi-supervised learning.
Supervised, Unsupervised, and Semi-Supervised
Supervised learning is where you have the data and the label. You are then telling the computer, this is data and this is the label now build a model.
Unsupervised learning is looking for structure in the data. All you have is the data and there is no label. So, you take the data and pass it through the unsupervised learning algorithm. Then it spits out and says this is a structure I have found. These sets of data belong to the same group. At the end of the day, you now start making sense of what those groups are. That is an example of unsupervised learning.
Next, semi-supervised learning involves some of the data you have labeled and some are not. So that is where the semi-supervised is because you have access to some of the data points.
Another Field of AI Machine Learning
But, reinforcement learning is another field of AI that is gaining a lot of traction recently. Especially with DeepMind and other companies that are doing good with reinforcement learning. The root of reinforcement learning is the Markov Decision Process. Not to go into too much detail, but reinforcement learning is like the science of decision making. From this state I have now, what action would I take from the current state that I have, and what reward will I get?
So reinforcement learning is a state, action, and reward process. It is unlike machine learning where you have data and you have labels. Most often reinforcement learning is an online learning process.
What The Future Holds
The AI algorithm does not know what the future is going to be. The answer in the future is going to learn as the time stamp increases. As far as I’m concerned it is important that there is a fusion between the branches of AI.
For example, in Q-learning, you have principle-based algorithms combined with deep learning. With these two coming together, you have very good results with Q-learning. The DeepMind’s AlphaGo has Q-learning at its root in defining and understanding. Q-learning is making value, computing value functions and more things around reinforcement learning.
What AI Machine Learning Is Doing In Society
So humans, as I said earlier, we’re very good at making some decisions. But we are very bad at computing when the sample or state space is large. Then we start making the wrong decisions. But with the right machine learning algorithms, we can compute large state spaces. This makes it so that we can make informed decisions.
Recently, we have algorithms that are doing very well on Jeopardy. There is an AI machine that was recently involved in a debate. Facebook’s AI is playing card games now. We have DeepMind’s AI playing video games and beating all the champions. This is because of the fusion of different algorithms. Researchers are trying many combinations and this is bringing about good results. The future is bright as far as I am concerned in this field.
There are going to be a lot more interesting things in the next few years. Researchers are sharing ideas and the fields coming together to make things happen. The simplest way to define deep learning is as a branch of AI that is trying to mimic the human brain.
Understanding How It Works
We humans, we have all these billions, hundreds of billions of neurons in our brain. Each neuron has what looks like an activation function. So the dendrites take in signals from whatever nerve is in our body. It gets to a junction there, and then sends that, and there is a pause. It’s an activation. It switches on or switches off. So deep neural networks try to mimic that. Then we have what looks like a single neuron. That neuron is a mathematical function, right there in the middle. That we pass in data like the neurons in our brain.
We pass in data into that neuron, and the neuron computes that data via an activation function. The activation function is the neuron’s house. There are so many activation functions around each of the neurons in deep learning. There is a sigmoid function and a rectified linear function. A lot of functions try to mimic what the brain does.
How It Relates To The Human Brain
So, like we have all these billions of neurons in our brain, we have all these billions of artificial neurons. These are all mathematical functions with linear functions. Also, they all have an activation to make it non-linear. If we combine this we tend to want to have the same effect as the neurons in our brains do. It is not one neuron. We have layers of neurons. So we have a hundred in one set of neurons. There are another hundred sets of neurons. So we pass in the input.
One layer computes the activation functions and gets the outputs. Then passes it into another set of neurons and passes into another set of neurons. That is where the deep comes in. So how deep can you go? Many tricks have made it easier by propagation to build deep neural networks. 150 layers deep.
Deep Neural Networks In AI Machine Learning
GoogleNet was one of the earlier versions of deep neural networks. They did well with image classification. But the thing is if it is too deep the gradients start to explode and diminish. So there is a trade-off between how deep you are able to go before your results start to become bad.
There are many tricks and researchers are trying out Inception networks as well. This network helps make it a lot deeper. Now you can take the activations from earlier layers and pass it into deeper layers. Then you can now build deeper neural networks.
To summarize the deep neural networks are artificial networks based on mathematical functions. They try to mimic the way the neurons in the brain behave. We don’t completely understand exactly how the neurons in the brain behave, but we have a rough idea. This has made it possible to build these artificial neurons that are doing pretty well in the real world.