It was her first writing course in college. Kate was given the task to create a 500-word essay on anything she wanted. Like most college students, she started the assignment exactly the night before it was due.
Long story short, Kate wrote something and received praise for it. This was the paraphrased interaction:
Professor: Let’s discuss this extraordinary essay.
Kate sits at her desk, smug.
Prof: It’s good, but obviously not because of the writing.
Kate sits at her desk, less smug.
Prof: It’s good because she was able to ‘see' what she wrote about.
And that’s it. Kate has some solid advice for people to see in order to improve their writing craft. Interestingly enough one of the suggestions reads:
“Look at the world around. Pay attention to details. Open your heart to what you see.”
What is seeing? Something physical. Our visual cortex is activated in a specific way that gives us information about the world around us.
But what is ‘seeing,’ like in Kate’s story? It’s different than in the physical sense. Another way to understand the ability to ‘see’ is ‘experiencing a thing for itself’. German philosopher Heidegger called it Dasein, as in “being there,” or “being in the world.”
Phenomenologies, Dietrich von Hildebrand, uses the language of a datum to describe our experience with reality: “We must seek to analyze the datum, delve into its nature, explore its relations with other fundamental data of experience, and finally inquire into the presuppositions that have to be fulfilled.”
The word machine learning was first used by Arthur Samuel. Samuel defined machine learning as “the field of study that gives computers the ability to learn without being explicitly programmed.” Later Tom Mitchell clarified that machine learning is when computers learn from experience to do some task.
The goal of AI has always been unclear – to imitate humans or perfect us? In either case, AI functions as a system of digesting experience and obtaining meaning from it. Experience takes the form of data. Data can be anything: photos, sounds, words, videos, and more. Machine learning finds patterns within large amounts of data. Large amounts of data aren’t absolutely necessary for machine learning to work but as mentioned in a previous post on the history of AI: the more data, the better the results.
There are a few different ways an algorithm can learn. One is called supervised learning. Supervised learning works by inferring a mathematical function from labeled data. To use labeled data, the data must be trained first. This training process involves a human to sift through the data and decide how to label it.
For example, let’s assume you were trying to train a neural network to identify a cat in a photo. After collecting images of cats in photos, someone would have to go through these images and label if the photo contains a cat, or not.
The supervised algorithm takes this labeled training data and then finds the inferred function. The inferred function becomes the guide for a neural network; the neural network uses this function to guess at new examples.
If there are certain rules that the person followed who initially labeled the data, then the neural network will guess these trends. For instance, let’s say that the person labeling the data did not label photos as a ‘cat’ if the photo contained a cat without a tail. The trained data would infer this trend, and when someone would like to look for cats using a search engine that is powered by this particular neural network, theoretically no photos with tail-less cats would show up.
Besides the traditional classification tasks, there are other methods powered by trained data, such as segmentation analysis and object detection. This section of the frontier is pushed forward via the data labeling process. Supervised learning has made substantial strides in the furthering of the frontier and with amazingly accurate results.
There are other methods of training, including unsupervised machine learning and reinforcement learning.
Unsupervised learning also infers a mathematical function from data, but the data doesn’t have to be unlabeled. There is no need to involve humans to train your data. However, the accuracy of the resulting function is lower than with supervised learning. Training a neural network with unsupervised learning typically requires an extremely large dataset.
Reinforcement learning trains a neural network through the use of a punishment and reward system. This type of training also requires trained data.
The process of machine learning involves collecting data, training the model with the data, and testing the model. If the model works well, we keep using it and improving the model with future data. If various fine-tuning and adjustments don’t work, we choose a new model and begin again.
Deep learning goes deeper. Deep learning is a class of machine learning algorithms that use multiple layers to extract higher-level features from the data. Deep learning is essentially machine learning but “on steroids: it uses a technique that gives machines the enhanced ability to find—and amplify—even the smallest patterns.” Let’s consider an image of a cat. One layer might be able to identify the edges of the cat’s body. Other, high levels, might identify ears, whiskers, or eyes. Deep learning is a powerful tool because the techniques can recognize even the most minute of patterns.
Andrew Ng did exactly that in 2012. Ng trained a neural network to recognize a cat with data from 10 million YouTube videos. Deep learning enabled engineers to find practical applications for machine learning and push the boundaries of the field even further.
Machine learning is a way for computers to take in experience and learn from it just like humans do. And the results have been phenomenal. Will machines outperform humans? It’s already been done with some tasks. It’s unclear if AI will ever be able to experience the world in the same way that I do. Even so, there’s a long road ahead and we’re just getting started.