A Snapshot of AI, Machine Learning, and Deep Learning

03.07.2018

Machine Learning

Real-world examples, prominent use cases, and AI’s (artificial intelligence) potential

The industrialized world is full of machines that outdo humans in strength and speed. Cranes lift steel beams to towering heights, car engines send passengers flying down the road at impossible speeds, and tree shredders pulverize entire pine trunks in a snap. Such inventions replicate and vastly outperform humans at tasks requiring physical exertion. They possess “artificial brawn”.

What then is artificial intelligence?

At a first pass, we can think of AI as a machine (i.e., a computer) replicating human cognitive tasks. A calculator, for example, embodies basic AI to do a job for which humans must use their brains: math.

While definitions of “artificial intelligence” abound, Peter Norvig and Stuart Russell, forefathers of the discipline, explain it well when talking about “rational agents”. A rational agent, they say, “acts so as to achieve the best outcome or, when there is uncertainty, the best expected outcome”. Most explorations into AI fit the “rational agent” description in some way.

How do “best” and “best expected” outcomes differ? A best outcome can be proven to be so. But with best expected outcomes, you cannot deduce that one answer, solution, or action is superior to another. For example, let’s say someone visits a golf retailer website and asks the chatbot on the site, “When does your San Francisco location close today?” In this case, there is only one right answer: 9 pm. We can verify that the chatbot made the best decision.

But if the person says, “I broke one of my clubs today. What should I do?”, it is unclear how the chatbot should respond. A rational chatbot would, in this case, leverage data from previous customer interactions to ask appropriate clarifying questions.

How Do Conventional AI Systems Work?

Conventional AI systems are pre-programmed, or “hard coded,” to solve specific problems. Developers write a staggering number of “if-then” logic statements into the algorithms. An AI system can only do what it’s explicitly programmed to do, its range of skills can be quite narrow. Not to mention, it demands colossal amounts of human labor to build.

Certain chess-playing computers are hard coded, for example. In this scenario, a data scientist might collaborate with a skilled chess player and enter individual chess moves into the AI system. By obeying these pre-defined commands, an AI agent can do rather well at the game. But it cannot do much else. If you asked it who Nelson Mandela is, it would fall silent.

To match a human being’s intelligence, an AI agent must acquire general knowledge and navigate multiple domains. Machine learning is the branch of AI that attempts to endow computers with this broader skill set.

What Is Machine Learning?

In machine learning, the data scientist, rather than dictating the AI system’s every action through toilsome pre-programming, creates a “learning algorithm” that teaches the AI agent how to, well, learn.

Also called a model, a machine learning algorithm must be trained with vast quantities of data. Teaching a model to play chess, for example, involves feeding it thousands of videos of different chess moves. Over time, the model learns to play the game – often at an elite level – by detecting patterns in the data.

Machine learning gives us things like spam filters, personalized recommendations on websites, and automated identity verification services. It also enables us to predict with stunning accuracy when our product inventory will run low or power will go out or prices will fall.

Broadly speaking, there are two ways to train a machine learning model: with or without supervision.

In supervised learning, you label the training data. Put another way, you give the model an answer sheet. For example, if you were to train an AI agent to identify a car versus an airplane using only text data, you would first compile and label the different “features” of your data. A feature set in the data may include the speed (feature 1) and weight (feature 2) of the vehicle. Next, you would glut your AI model with — ideally — thousands of feature sets. Eventually, the algorithm would catch on to the drastic speed and weight disparities between the two vehicle types and learn to recognize cars and planes on its own.

As you might’ve guessed, the cheat sheet is absent in unsupervised learning. While still exposing the algorithm to feature sets, you leave out the solutions. In this context, you might program the model to seek out specific outputs as it combs through the feature sets or tell it to “cluster” data according to patterns it finds. Scientists leverage this method in genomic sequencing. They feed machine learning models enormous, unlabeled data sets containing genetic information and allow the model to gather certain DNA into clusters based on shared traits.

Genomic sequencing turns out to require a lot more computing power and analytical depth than traditional machine learning use cases. For such a massive undertaking, researchers employ an advanced form of machine learning called deep learning.

What Is Deep Learning?

Deep learning draws upon reams of big data, huge reservoirs of computing power, and sophisticated machine learning techniques to notch some of AI’s greatest achievements to date – Google’s deep learning agent AlphaGo beating the world champion of Go, a complex board game, being just one example.

The algorithms of a deep learning model are arranged in a neural network, a signature feature of deep learning. Neural networks (nets) mimic the process of the human brain in which certain neural connections grow stronger as the brain responds to stimuli.

The neurons of deep learning are called nodes, and they occupy multiple layers within a model. A node in the first layer will bond with a node(s) in the second layer and so on. As this happens, each node connection gets assigned a weight, which, when aggregated, largely determines the model’s output.

Neural nets are much easier to understand in real-world examples. Let’s say, for example, we want to train a deep learning model to recognize Victorian houses within photographs – a skill known as image classification. When you show the model a photo, nodes in the first layer begin to identify “simple features” in the image such as a color, a texture, or a flat edge. These frontline nodes will then relay what they have learned to their counterparts in the next layer. Weights are assigned to each nodal connection at the same time.

Deep Learning

The picture becomes clearer as information is volleyed across more and more layers. At a certain point, the model may begin to identify a slanted bay window in the photo, for instance. Finally, when nodes in every layer have dissected the photo, and all the weights have been added up, the deep learning model will provide its answer (output). It may declare that it is 95% sure it sees a Victorian house, 3% sure it sees a Craftsman, and 2% sure it sees an adobe. Computer scientists will then adjust the weight values to improve the model’s accuracy for future iterations.

Thankfully, the utility of deep learning is not confined to bygone architecture. Deep learning methods have been employed in healthcare, for example, to spot cancerous cells in human tissue. Some AI agents have even excelled physicians in this area.

Beyond medical imaging, deep learning models are used to process all sorts of data — not just photographs — and are fueling advancements in cyber fraud protection, pharmaceutical research, self-driving cars, predictive analytics, and more.

Why High-Quality Training Data Is So Vital to AI’s Development

Modern AI systems need mountains of high-quality data to learn how to behave rationally. Were it not for the rise of Big Data, we would still be groping our way toward rudimentary machine learning milestones. But procuring rich data is a challenge for many researchers and data scientists. Even if they have thousands of feature sets, they may lack the time and staff to label the data.

Thankfully, clickworker, a provider of AI training data via crowdsourcing, can supply you with heaps of clean, high-quality data. clickworker offers personalized training data, including text conversations, web research, videos, photos, and voice recordings. Whether you’re developing an AI model to do natural language processing, predictive analytics, image classification, or something else, you can innovate faster and at a favorable cost with clickworker’s custom data sets.

AI Training Data Service at clickworker
Click on this link “AI Training Data by clickworker” or call clickworker directly (+49 201 9597180 or +1 415 6897781) to learn more about the service.

The next wave of AI breakthroughs cannot come soon enough in some cases. Indeed, the hope of artificial intelligence is that it will lift great burdens from the shoulders of humankind and help solve our most perplexing challenges. If we are to reach this promised land, we are going to need a lot of training data. Of that we can be sure.

 

Autor: Matthew D.