Linear regression is one of the simplest statistical models, but don’t let that fool you. It is a powerful tool because of that simplicity. Understanding linear regression starts with the name. A regression analysis is simply a method of estimating the relationship between a dependent variable and a set of independent variables. For example, given a set of data about childhood measurements like height, weight, gender, and so on, can we estimate the relationship between those variables and the dependent variable, which is the child’s height as an adult?

We can see the result of that regression plotted along with the data. The line is our model’s prediction of petal length for each sepal length. Of course, there will be some error because there is some variability in the data, but the fit generally describes the relationship well. The equation for this relationship is:

We’ve seen how we can calculate the terms of a linear model using minimization of a residual equation, but this is not the end of the model building process. Unlike machine learning models, which have millions of parameters obscured in a network, each parameter of a linear regression model can be interrogated and analyzed. The regression model produces several metrics for use. First, there is the R² measure or the Coefficient of Determination. R² measures how much of the variance in the data is explained by the model. A value of 1 means that the model perfectly describes the data while a value of .5 suggests that only 50% of the variance across the data is explained in the model. The rest is unaccounted for. Obviously, the closer to 1 the R² value, the better.

In the past few years, more and more organizations have started relying solely on Machine Learning (also called Artificial Intelligence or AI) for addressing Text Analytics tasks. The reasons are that this approach is faster, suitable for large volumes of data, and more adaptable to new contexts, as it does not require hard-coded rules. However, AI / Machine Learning algorithms are only as good as the data they are trained on.

What about the Rule-Based approach?

Even though it has been abandoned lately in favor of machine learning solutions, it offers certain advantages over the latter method. Since the rules are created by humans, the results are consistent and more accurate. It is also easier to understand the logic behind the machine’s decisions and improve a model, instead of trying to comprehend a “black box”, which is the case with most machine learning algorithms. But the main issue with this approach is that the rules are hard-coded and are limited to the existing context, making it hard to adapt to new context; this results in higher precision, but lower recall.

Thus, with a combined approach, we achieve both the high precision that Rule-Based approaches bestow on created training datasets, along with the high recall that is achieved through Machine Learning in different contexts. If you are interested in automating the generation of accurate training data to improve the results of your Machine Learning models, contact us and learn about the tools we have available.

Model frameworks have a set of qualitative measures attached to them which, while difficult to directly measure, can be useful for thinking of. We will consider two qualities: capacity and complexity. Capacity is the measure of a model framework’s ability to describe data distributions; it can be thought of as potential accuracy. For example, imagine we scatter pebbles on the ground and would like to model the shape they form with two different models. One model is a rigid stick (depicted by the blue line) and the other is an elastic band (depicted by the red line).

The stick has very low capacity because it can only model one shape while changing the angle that it lies on the ground. A few pebbles may fit into this model, but it is unlikely to be a good choice for many situations. The elastic band, while probably not being perfect, can twist and bend to create a host of complex curves to fit our data. This model can fit more potential data distributions and has a higher capacity.

Capacity is a very important property of models. If our model’s capacity is too low, we may never be able to adequately make predictions no matter how much training data we use or how fancy our hardware is. Models with high capacity include Neural Networks and Support Vector Machines (SVMs), and this is why they have been so popular.

However, capacity is not the only property to consider. Another quality is complexity, which can also be thought of as interpretability or how easy it is for a human to understand. In our previous example, the stick had low complexity (or high interpretability). It can be described and understood well by humans, and it is highly generalizable. Low complexity models include Linear Regression and Decision Trees. Conversely, the many curves created by the elastic band are very complex and difficult for humans to describe. SVMs and Neural Networks would, therefore, be considered high complexity models.

This relationship between complexity and capacity that we saw in the above example is generally true for all of our models. By increasing our capacity, we often must incur the cost of increased complexity. This is part of the analysis you must consider when choosing a model. Do you need to be able to describe or personally understand what the model is doing? If so, going all in for the most complex but accurate model may not be desirable. This is why many hospitals and healthcare-related institutions often rely on less complex model structures like Random Forests. When taking the health of patients into account, doctors want to be able to understand what a model is doing before taking action on those results.

Resources

Thus far we have discussed the theoretical constraints of a trained model. Now we should consider more practical properties. A machine learning model requires training, which in turn requires data, processing power, and time. When selecting a machine learning model to use, we need to weigh how much of these resources we have available to invest in the model. Additionally, the quantification of the model itself is important. For example, if we plan to deploy the model on mobile technology, we should be cautious about how much storage the model itself requires. As expected, if we desire models with high capacity, we usually need to invest more resources into the model during training and storage. Let’s consider a sample of models and discuss the resources they require.

On the low-cost end of the spectrum are Naïve Bayes models. These models are generally lower in Capacity but are extremely easy to interpret and train. One of the main advantages of Naïve Bayes models is that they are exceedingly fast to train and store. Unfortunately, the model relies on the Naïve Bayes assumption which is rarely true for our data. However, if your data does not stray too far from that assumption, you can find that the model provides decent results in a cheap package.

Also relatively low-cost are Decision Trees and Random Forests. These models are interpretable and the tree structure of the models make them easy to store efficiently if you don’t have an enormous number of features. Unfortunately, we start to see an increase in the time required to train for these models.

SVMs, a popular technique, come with a large jump in resources required. Training these usually requires more data than other methods and the resulting models are bulky, almost impossible to interpret into plain English, and tricky to train in the first place because of the choice of kernel. However, SVM’s are powerful models when trained correctly, and thus they are widely used.

Possibly the most costly of the models are Neural Networks. Famed for their high capacity, these models require extensive processing and time to train and can take up a large chunk of storage to contain. But it isn’t all doom and gloom about the resources required for Neural Networks. Because of how they are trained, Neural Networks can take additional data to finetune the training at any point, meaning that we don’t need to retrain our models every time we get new data.

Which Model to Pick?

Ultimately, the answer is not obvious in every case. When choosing which machine learning model to use, we need to consider many different factors: What tradeoff do we want to make between capacity and complexity? What resources do we have available for our training? Do we want to be able to store the model in a compact manner? Do we want to be able to continually use new data without having to retrain the model entirely each time? Finally, we don’t always make the right choice. Sometimes we have to experiment with multiple structures to find one that fits our needs. If you can, train multiple models to help decide which one to devote your attention to.

There are many reasons for the explosion of machine learning advancements over the past decade. We now have vastly improved hardware for fast computation, and memory is cheaper than ever. Data is now “Big Data,” and it is both jealously hoarded and publicly available in repositories such as ImageNet. Individually, these advancements are already a blessing for the technology-space. But for artificial intelligence (AI), they have opened the gates for something truly powerful—Neural Networks.

If that all sounded magical and farfetched, then don’t worry—it is. How exactly are we supposed to arrange these Neurons in such a perfect balance that their combined minuscule computations lead to a machine recognizing human faces? We can’t. So how do we get this to work? Well, we are talking about Machine Learning after all. And Machine Learning is how we are going to solve this. We aren’t going to calibrate the Neurons to be in balance. They are going to calibrate themselves.

In order to achieve this, we need training data. We need data that is labeled with the desired output we want from the machine. We can provide this data to the unconfigured Neural Network. It will process the data and likely output something nonsensical and useless. But this is fine. We can use a mathematical function called Error. This Error is just a measurement of how different our output was from our desired targets. Then, using Calculus,we can discover how much of that Error is caused by the calibration of each one of our Neurons! Using this information, we can then tune the Neurons slightly and repeat the process—feed data into the Network, observe the output, calculate the Error, and use Calculus to know how to tune the Neurons to make them more accurate. This process continues over time until we have converged to a calibrated Network. This process of using Error functions, Calculus, and tuning is what Machine Learning is.

Power at a Price

Although it is conceptually simple (at least in this explanation that ignores the details), it is very resource intensive. At the moment I am fine tuning a Neural Network on my own desktop to recognize Western Art styles. Although the dataset is only a few GB, it takes almost an hour to run a single iteration of the training. It will take nearly two days to complete the 50-iteration learning schedule I planned. And even then, I may need to schedule another one if the Network still needs to learn more! And I’m not running this on some dusty machine I dragged out of the aughts. This is an almost brand-new desktop powered with a 3.7 GHz AMD Rhyzen 8-core processor with 32GB of RAM available and virtually no other load on the machine.

Neural Networks are expensive to train. If you want to increase performance, you could pay for an expensive GPU, but that might set you back nearly a thousand dollars. Companies and research institutions may have the funds to throw at this problem, but individuals, small companies, and small research groups may not. Luckily, they don’t need to anymore.

Cloud services have opened access to remote processing to provide other computing options to those desiring to train a Neural Network. Don’t have an expensive rig? No worries, just rent one remotely from Google at a fraction of the price.

Neurally Networked World

Neural Networks are here and they aren’t leaving. New advancements and computing architectures are constantly being published. Convolutional Networks are good at processing images and Recurrent Networks can handle variable sized data, streaming data, or sequential data. Neural Networks are powerful, indeed—far more so than other solutions we have. But they don’t mimic what is really occurring in our brains. There are some deep flaws even in our most advanced networks. For instance, it is extremely easy to confuse a Neural Network. A Network designed to recognize stop signs can be fooled by a few well-placed stickers on a sign. There are deep safety concerns if these are going to be used in self-driving cars for example.

Neural Networks are deeply dependent on the data used to train them (just as we discussed in a previous article). They are also dependent on how we configure their output. Most networks are designed to give a decision. For object detection, the network must return what it thinks the input image is. But what happens when we feed the network “nothing,” such as a completely blank image or fuzzy static noise? The network is forced to return something so it “sees,” say, a dog in the empty space. This is nonsense. Why would it choose one object over another in these cases where a human would just refuse to rigidly define nonsense? As another example, if we slightly alter an image by inserting some imperceptibly small random noise, we can completely trick a Neural Network. Where it before correctly thought that the image was a butterfly, it now thinks the image is a truck, while a human sees no difference in the images. This is bad, and it reflects deep issues with Neural Networks as we construct them today.

Neural Networks are occupying a liminal space. They are simultaneously scarily powerful and laughably simple and ignorant. They can best the human masters and yet be duped by the smallest of changes. We won’t be seeing Neural Networks achieve human-like sentience any time soon, and they aren’t ready for deployment in many other types of systems. But they are already being used in ways that should cause alarm.

China is already using facial recognition technology to tag members of the Uighur ethnic minority group.  Accurate voice recognition ensures individuals could potentially be tracked even if they are not near a camera by turning the phones in our pockets into monitoring devices. “Deep Fakes” are a growing type of video that can modify existing videos to map one person’s face and voice over another’s to create a fake video that could be used for blackmail or disinformation. While Terminator remains science-fantasy, armed military drones using neural networks for stabilization, navigation, targeting, and tactics would revolutionize armed conflicts in ways impossible to predict.

Though neural networks are being used to oppress in some places they are also being used to save in others. Medical facilities increasingly deploy network systems to detect ailments such as cancer or infectious diseases. Laboratories can use similar networks for modeling complex biomolecules and developing treatments. Neural networks are even being used in traffic light control systems to increase vehicle flow and reduce accidents.

Where will Neural Networks take us next? It’s hard to say. But it seems evident that the world is caught and will remain in a web of Neurons.

To demonstrate how these options work, let’s imagine we are working with a dataset that has records (or rows) of textual responses that are either in English or Spanish. Regardless of whether we want to use machine translation or to process individual documents in their native languages, we first need to split the dataset up by language. In this example, all the English records are together in one dataset, and all the Spanish records are in another. We then use the software PolyAnalyst for Text™ to perform native language text analysis by implementing the corresponding language modules. Currently, this software can process the texts of 16 different languages, and it can integrate with third-party translation services such as Microsoft, SDL, and Google. Splitting the data can be easily accomplished by using a combination of the Language Detection and Filter Rows nodes in PolyAnalyst. The Language Detection node automatically samples the text of each record and determines what language is being used (as seen below). Once each record is tagged with its language, we can use these tags to filter out the data by language.