Not very long ago, sometime towards the end of the first decade of the 21st century, internet users everywhere around the world began seeing fidelity tests while logging onto websites. You were shown an image of a text, with one word or usually two, and you had to type the words correctly to be able to proceed further. This was their way of identifying that you were, in fact, human, and not a line of code trying to worm its way through to extract sensitive information from said website. While it was true, this wasn’t the whole story.
Turns out, only one of the two Captcha words shown to you were part of the test, and the other was an image of a word taken from an as yet non-transcribed book. And you, along with millions of unsuspecting users worldwide, contributed to the digitization of the entire Google Books archive by 2011. Another use case of such an endeavor was to train AI in Optical Character Recognition (OCR), the result of which is today’s Google Lens, besides other products.
Do you really need millions of users to build an AI? How exactly was all this transcribed data used to make a machine understand paragraphs, lines, and individual words? And what about companies that are not as big as Google – can they dream of building their own smart bot? This article will answer all these questions by explaining the role of datasets in artificial intelligence and machine learning.
ML and AI – smart tools to build smarter computers
In our efforts to make computers intelligent – teach them to find answers to problems without being explicitly programmed for every single need – we had to learn new computational techniques. They were already well endowed with multiple superhuman abilities: computers were superior calculators, so we taught them how to do math; we taught them language, and they were able to spell and even say “dog”; they were huge reservoirs of memory, hence we used them to store gigabytes of documents, pictures, and video; we created GPUs and they let us manipulate visual graphics in games and movies. What we wanted now was for the computer to help us spot a dog in a picture full of animals, go through its memory to identify and label the particular breed among thousands of possibilities, and finally morph the dog to give it the head of a lion that I captured on my last safari. This isn’t an exaggerated reality – FaceApp today shows you an older version of yourself by going through more or less the same steps.
For this, we needed to develop better programs that would let them learn how to find answers and not just be glorified calculators – the beginning of artificial intelligence. This need gave rise to several models in Machine Learning, which can be understood as tools that enhanced computers into thinking systems (loosely).
Machine Learning Models
Machine Learning is a field which explores the development of algorithms that can learn from data and then use that learning to predict outcomes. There are primarily three categories that ML models are divided into:
These algorithms are provided data as example inputs and desired outputs. The goal is to generate a function that maps the inputs to outputs with the most optimal settings that result in the highest accuracy.
There are no desired outputs. The model is programmed to identify its own structure in the given input data.
The algorithm is given a goal or target condition to meet and it is left to its devices to learn by trial and error. It uses past results to inform itself about both optimal and detrimental paths and charts the best path to the desired endgame result.
In each of these philosophies, the algorithm is designed for a generic learning process and exposed to data or a problem. In essence, the written program only teaches a wholesome approach to the problem and the algorithm learns the best way to solve it.
Based on the kind of problem-solving approach, we have the following major machine learning models being used today:
These are statistical models applicable to numeric data to find out a relationship between the given input and desired output. They fall under supervised machine learning. The model tries to find coefficients that best fit the relationship between the two varying conditions. Success is defined by having as little noise and redundancy in the output as possible.
Examples: Linear regression, polynomial regression, etc.
These models predict or explain one outcome among a few possible class values. They are another type of supervised ML model. Essentially, they classify the given data as belonging to one type or ending up as one output.
Examples: Logistic regression, decision trees, random forests, etc.
- Decision Trees and Random Forests
A decision tree is based on numerous binary nodes with a Yes/No decision marker at each. Random forests are made of decision trees, where accurate outputs are obtained by processing multiple decision trees and results combined.
- Naïve Bayes Classifiers
These are a family of probabilistic classifiers that use Bayes’ theorem in the decision rule. The input features are assumed to be independent, hence the name naïve. The model is highly scalable and competitive when compared to advanced models.
Clustering models are a part of unsupervised machine learning. They are not given any desired output but identify clusters or groups based on shared characteristics. Usually, the output is verified using visualizations.
Examples: K-means, DBSCAN, mean shift clustering, etc.
- Dimensionality Reduction
In these models, the algorithm identifies the least important data from the given set. Based on the required output criteria, some information is labeled redundant or unimportant for the desired analysis. For huge datasets, this is an invaluable ability to have a manageable analysis size.
Examples: Principal component analysis, t-stochastic neighbor embedding, etc.
- Neural Networks and Deep Learning
One of the most widely used models in AI and ML today, neural networks are designed to capture numerous patterns in the input dataset. This is achieved by imitating the neural structure of the human brain, with each node representing a neuron. Every node is given activation functions with weights that determine its interaction with its neighbors and adjusted with each calculation. The model has an input layer, hidden layers with neurons, and an output layer. It is called deep learning when many hidden layers are encapsulating a wide variety of architectures that can be implemented. ML using deep neural networks requires a lot of data and high computational power. The results are without a doubt the most accurate, and they have been very successful in processing images, language, audio, and videos.
There is no single ML model that offers solutions to all AI requirements. Each problem has its own distinct challenges, and knowledge of the workings behind each model is mandatory to be able to use them efficiently. For example, regression models are best suited for forecasting data and for risk assessment. Clustering modes in handwriting recognition and image recognition, decision trees to understand patterns and identify disease trends, naïve Bayes classifier for sentiment analysis, ranking websites and documents, deep neural networks models in computer vision, natural language processing, and financial markets, etc. are more such use cases.
The need for training data in ML models
Any machine learning model that we choose needs data to train its algorithm on. Without training data, all the algorithm understands is how to approach the given problem, and without proper calibration, so to speak, the results won’t be accurate enough. Before training, the model is just a theorist, without the fine-tuning to its settings necessary to start working as a usable tool.
While using datasets to teach the model, training data needs to be of a large size and high quality. All of AI’s learning happens only through this data. So it makes sense to have as big a dataset as is required to include variety, subtlety, and nuance that makes the model viable for practical use. Simple models designed to solve straight-forward problems might not require a humongous dataset, but most deep learning algorithms have their architecture coded to facilitate a deep simulation of real-world features.
The other major factor to consider while building or using training data is the quality of labeling or annotation. If you’re trying to teach a bot to speak the human language or write in it, it’s not just enough to have millions of lines of dialogue or script. What really makes the difference is readability, accurate meaning, effective use of language, recall, etc. Similarly, if you are building a system to identify emotion from facial images, the training data needs to have high accuracy in labeling corners of eyes and eyebrows, edges of the mouth, the tip of the nose and textures for facial muscles. High-quality training data also makes it faster to train your model accurately. Required volumes can be significantly reduced, saving time, effort (more on this shortly) and money.
Datasets are also used to test the results of training. Model predictions are compared to testing data values to determine the accuracy achieved until then. Datasets are quite central to building AI – your model is only as good as the quality of your training data.
With heavy requirements in quantity and quality, it is clear that getting your hands on reliable datasets is not an easy task. You need bespoke datasets that match your exact requirements. The best training data is tailored for the complexity of the ask as opposed to being the best-fit choice from a list of options. Being able to build a completely adaptive and curated dataset is invaluable for businesses developing artificial intelligence.
On the contrary, having a repository of several generic datasets is more beneficial for a business selling training data. There are also plenty of open-source datasets available online for different categories of training data. MNIST, ImageNet, CIFAR provide images. For text datasets, one can use WordNet, WikiText, Yelp Open Dataset, etc. Datasets for facial images, videos, sentiment analysis, graphs and networks, speech, music, and even government stats are all easily found on the web.
Another option to build datasets is to scrape websites. For example, one can take customer reviews off e-commerce websites to train classification models for sentiment analysis use cases. Images can be downloaded en masse as well. Such data needs further processing before it can be used to train ML models. You will have to clean this data to remove duplicates, or to identify unrelated or poor-quality data.
Irrespective of the method of procurement, a vigilant developer is always likely to place their bets on something personalized for their product that can address specific needs. The most ideal solutions are those that are painstakingly built from scratch with high levels of precision and accuracy with the ability to scale. The last bit cannot be underestimated – AI and ML have an equally important volume side to their success conditions.
Coming back to Google, what are they doing lately with their ingenious crowd-sourcing model? We don’t see a lot of captcha text anymore. As fidelity tests, web users are now annotating images to identify patterns and symbols. All the traffic lights, trucks, buses and road crossings that you mark today are innocuously building training data to develop their latest tech for self-driving cars. The question is, what’s next for AI and how can we leverage human effort that is central to realizing machine intelligence through training datasets?