Types of Machine Learning

If the distinction between artificial intelligence, machine learning, and deep learning wasn't already confusing enough, machine learning algorithms can be broken up into three types, which can also be combined to form hybrid learning algorithms:

1. Supervised learning
2. Unsupervised learning
3. Reinforcement learning

Supervised Learning. Supervised learning is the most straightforward learning problem to understand. It aims to mimic a teacher–student relationship through labeled data. In supervised learning, you provide one or more examples of the output expected for given inputs. Through the process of training, the model learns to generate results that are most similar to the training data. Supervised learning further breaks down into regression (predicting a quantity) and classification (predicting a label). While supervised learning is likely the most common form of machine learning, it isn't the best to utilize the data present with Big Data because the data often is unlabeled. In order to know the expected output, data must be labeled, which can be a time-consuming and expensive process. Advances in few-shot learning help alleviate this problem and despite the high initial cost of supervised learning, it can be hugely beneficial. Detecting product defects or safety violations using models can easily save thousands of dollars. Once developed, these models can often perform at a similar level to human evaluation, but scale much more easily.

Unsupervised Learning. On the opposite side, unsupervised learning requires no labeled data, instead learning patterns directly from the data. While unsupervised learning is excellent at pattern recognition and anomaly detection, it can be more difficult to apply to problems. Still, unsupervised learning can utilize Big Data more efficiently without requiring human interaction. This means that unsupervised learning can offer a lower upfront cost, although it is not universally applicable. Commonly, unsupervised learning is used for clustering and dimensionality reduction.

Reinforcement Learning. Reinforcement learning relies on an interactive environment that an agent can explore. The algorithm tends to learn by trial and feedback. Most commonly this is used for robotics, although recently it has been used for scheduling and chip design as well. While developments in reinforcement learning may well drive Smart manufacturing in the future, reinforcement learning is less commonly applied to problems and may require more specialized knowledge to produce.

Different combinations of these techniques can form hybrid learning algorithms. Semi-supervised learning combines supervised and unsupervised learning to utilize unlabeled data to improve results from a small labeled data set. Recently, semi-supervised learning has driven improvements in computer vision by allowing a better internal representation of images created over a larger unlabeled data set. Self-supervised learning utilizes an unlabeled data set for supervised learning. Usually, the raw data set is the expected result, while some modified data serves as the input. One example of this is natural language processing, where the model is often taught to predict a word removed from a sentence.

Common Computer Vision Tasks

In computer vision, there are a number of commonly researched tasks that machine learning currently dominates. Each task focuses on producing different information given an image. While some tasks, such as barcode scanning or motion detection, don't require machine learning, many more complicated tasks are solely viable with machine learning. Simple tasks may still provide large productivity gains; for example, color detection is an easy computer vision task, but may be utilized to generate machine status information without interfacing with proprietary protocols. For many tasks, computer vision is not only cheaper than other ways of solving the issue, but is also more generalizable. Image classification is the task of assigning a label to an image. This may be used to identify a state; for example, a model may classify machines as in-use, waiting for repair, or available. This could be used to identify defective products where the location of the defect isn't important to the process. Alternatively, image classification may look to recognize one of several pre-known objects in an image. In this case, it is called object classification.

Exhibit 9.3 shows an example of image classification at work.⁶

If the location of an object is important, object classification can be combined with object localization to perform the object detection task. The goal of object detection is to determine both class and location of an object in an image, if one or more exists. This may be a simple problem like locating all the people in an image, or more complex like identifying different parts for counting as they pass on a conveyor belt. Object detection is often used alongside other computer vision techniques to determine the state of a given scene. For example, instead of building an image classifier to determine whether an employee is wearing the appropriate personal protective equipment, building an object detector to recognize hard hats, safety vests, and safety glasses would allow greater information about why the behavior is unsafe. In many cases, the last level of logic is based on either regulation or general knowledge, which should be directly coded rather than learned.

Exhibit 9.4 shows an example of object detection at work.⁷

If object detection can be seen as an advanced form of object classification, image segmentation is a more advanced form of object detection. Image segmentation aims to divide an image into multiple parts based on some attribute. It can be broken into two main forms: instance segmentation and object segmentation. Object segmentation involves segmentation based on classes. It might provide segmented areas for the conveyor belt, metallic objects, and plastic objects in a given image. Instance segmentation generates separate segments for each instance of an object. Rather than three segments in the previous example, each nut, bolt, and tray would be identified as a separate segmentation mask.

Exhibit 9.5 shows an example of instance segmentation and object segmentation.

Identification is the computer vision task involving identifying specific instances of an object, usually over time. One example of a common identification task is facial recognition. Object identification could be used to track a product and keep a record of which workstations it passed through. This might allow better tracing of products that fail quality control as well as insight into bottlenecks in the production process.

Although there are many more computer vision tasks with differing state-of-the-art models, a few other notable mentions include optical character recognition, pose estimation, depth mapping, and 3D reconstruction. Optical character recognition (OCR) could digitize paper workflows and allow seamless transitioning between digital and physical tracking systems. Pose estimation may be used for predicting and preventing human–robot collision by determining 3D orientation and the position of the robot, human, or both. Depth mapping is more prevalent in the world of self-driving, but it can provide a uniform replacement or augmentation for radar or lidar data. 3D reconstruction is fairly new but could allow the creation of 3D models, automatically speeding up prototyping and design stages.

Building a Model

Once you know what you want to build, the next step is creating a data set and model. This section addresses the process of building a supervised learning model for object detection, but the same process will apply elsewhere. The basic process involves selecting data, cleaning and annotating data, training or fine-tuning a model, evaluating the trained model, and continual improvement.

While not glamorous, selecting and cleaning the data are the most time-consuming and important steps in machine learning. Biases or faulty data can easily poison a model, reducing accuracy by a noticeable margin. At this stage, you should consider how much you are going to spend on data collection and annotation on a weekly, monthly, or yearly basis. Unfortunately, models often require regular retraining to handle issues like concept drift, where the statistical properties of a target change as the environment in which the model works changes. For example, if you build a model to detect safety vests, but the company changes the color of all safety vests, your model likely will perform worse because it hasn't seen the new color before. When selecting data, you want to make sure you get a representative set from all the different camera angles and conditions you expect the model to see when deployed. Ideally, there should also be a balance of classes that you expect to see, without any one class being over- or underrepresented. You should also make sure to be aware of any biases or shortcomings in the data set, preferably documenting them alongside the model. The siloed nature of data at larger companies can interfere with the ability to get widespread and useful data selected. Where possible, long-term investment in a single data lake can make this process faster, cheaper, and more agile.

After data selection and some analysis, cleaning and annotation of the data pose a number of unique challenges. Before starting annotations, you should write up a guide for how to annotate the data. This minimizes communication errors and allows more uniform results, which in turn leads to better model performance. While you may need to annotate the data internally, external annotation services likely offer better guarantees for annotation accuracy. Internally, you will likely have to train team members and double-check annotations, since everyone will have a different idea about annotating initially.

Once you have your annotated data, you should split it into three sets: training, validation, and testing. At least 10% of your data should be set aside for testing and around 20% should be set aside for validation. This leaves the remaining 70% for training. These data sets are kept separate to ensure that you have adequate data to estimate accuracy of your model. If there is overlap between these data sets, you will have no way of performing unbiased evaluation of your model or ensuring that no overfitting occurred. Overfitting is when the model performs better on the training data than on the test/validation data. This is usually because it conforms too closely to the data, effectively memorizing it rather than generalizing. You want to stop training before you overfit the data. The purpose of a separate validation set is to allow you to experiment with parameters of different models without selecting a model that best fits the test set. This becomes more important if you are testing a large range of models or parameters but is good practice either way.

Now that you have data, you can start experimenting with models. While state-of-the-art models may offer higher accuracy, they often are more expensive to train and deploy because they are more complex. One good way to evaluate model complexity is to look at inference time. The longer it takes to perform inference on one or many images, the more complex it is and the longer it will take to train. If you can, start with simple initial models such as MobileNet or an SSD-based model and scale up if you need better performance. You can fine-tune (train only the last few layers of a model) or train the entire model, although the latter requires significantly more data to do without overfitting. In general, fine-tuning will be enough for most purposes. Like model complexity, start simple and see whether you need more.

Once you have a trained model, you should check its performance on the validation set. This will allow you to see how well it performs on data it hasn't seen before. From here, repeating the training step with different models or different parameters can produce better results. When you are happy with the results, evaluating on the test data set should give an indication of the performance you will see for the model when deployed.

From here, the model can be updated with whatever frequency is necessary to obtain good performance. One of the greatest benefits of machine learning is that more data and computation tend to produce better results so there is a straightforward improvement process. Over time, you can build up a data set allowing a better model to be made. This data set can also be a valuable asset to any company, providing a benefit over your competitors.

Machine Learning Pipelines

Given the continuous nature of machine learning, it may make sense to have a pipeline, much like you would for data science. Pipelines may also offer cost-saving techniques such as active learning, which are beyond the scope of this chapter. Machine learning pipelines can easily become a greater time and maintenance sink than the models they were used to build.⁸ If your needs are met by one of the many existing free or paid pipelines, these will almost always offer more features than a pipeline you build yourself.

Having a pipeline can also be beneficial by allowing integration of automated tests to check whether the model is performing well. Organizing and tracking model performance can help diagnose issues and regressions before they cause major issues.

1. What is the relationship between artificial intelligence, machine learning, and deep learning?
   1. There is no relationship.
   2. Machine learning and deep learning are subsets of artificial intelligence, but completely separate from each other.
   3. Deep learning and artificial intelligence are subsets of machine learning.
   4. Artificial intelligence contains machine learning. Deep learning is a subset of machine learning.
2. Artificial intelligence experienced how many winters?
   1. Zero
   2. One
   3. Two
   4. Three
3. Which of the following have enabled the rise of machine learning?
   1. Big Data
   2. Moore's law
   3. Specialized hardware
   4. All of the above
4. Which model architecture revolutionized computer vision by outperforming other techniques?
   1. Faster R-CNN
   2. Convolutional Neural Networks
   3. AlexNet
   4. Reinforcement learning
5. What is the difference between supervised and unsupervised learning?
   1. Supervised learning requires constant human interaction; unsupervised learning solves more problems.
   2. Unsupervised learning works on data without labels, while supervised learning requires labels.
   3. Unsupervised learning requires an interactive environment which an agent explores, while supervised learning requires labeled data.
   4. All of the above
6. Which of the following is not an example of image classification?
   1. Finding a person in an image
   2. Checking whether an image is a cat or a dog
   3. Determining whether a machine in an image is on or off
   4. Classifying whether the image is taken on the East or West Coast.
7. What should you ensure that you do with training, validation, and testing data?
   1. Mix the data to ensure that you have an even distribution of each when training and evaluating your model.
   2. Keep the data separate and don't use testing data to evaluate different models.
   3. Train until the model performs significantly better on training data than validation or test data.
   4. Set aside at least 50% of your data for validation and testing.
8. When building an artificial intelligence system, you should always:
   1. Consult with stakeholders to determine potential consequences of the system.
   2. Develop and deploy systems quickly to determine whether they can accelerate your process.
   3. Test systems by implementing them in your manufacturing facility.
   4. Avoid artificial intelligence and machine learning at all costs.
9. Which of the following is an example of an adversarial attack?
   1. A sticker is placed on a machine to prevent it from properly detecting the machine state.
   2. A model fails to recognize a person in an image because of a change to safety vest color.
   3. A model learns to detect machine state by image brightness because the training data features machines in their “off” state at night.
   4. A cobweb obscures the camera's view and causes poor model performance.
10. Who is responsible for a machine learning algorithm that causes harm?
    1. Nobody; the algorithm works in obscure ways.
    2. It is difficult to say because there is little legal precedent, but the company or anyone involved in the deployment of the system may be liable.
    3. The person harmed by the algorithm.
    4. The company.