In this article, Yusuke Sugomori, the author of the book Deep Learning with Java, we’ll first see how deep learning is actually applied. Here, you will see that the actual cases where deep learning is utilized are still very few. But why aren't there many cases even though it is such an innovative method? What is the problem? Later on, we’ll think about the reasons. Furthermore, going forward we will also consider which fields we can apply deep learning to and will have the chance to apply deep learning and all the related areas of artificial intelligence.
The topics covered in this article include:
- The difficulties of turning deep learning models into practical applications
- The possible fields where deep learning can be applied, and ideas on how to approach these fields
We'll explore the potential of this big AI boom, which will lead to ideas and hints that you can utilize in deep learning for your research, business, and many sorts of activities.
(For more resources related to this topic, see here.)
The difficulties of deep learning
Deep learning has already got higher precision than humans in the image recognition field and has been applied to quite a lot of practical applications. Similarly, in the NLP field, many models have been researched. Then, how much deep learning is utilized in other fields? Surprisingly, there are still few fields where deep learning is successfully utilized. This is because deep learning is indeed innovative compared to past algorithms and definitely lets us take a big step towards materializing AI; however, it has some problems to be used for practical applications.
The first problem is that there are too many model parameters in deep learning algorithms. We didn't look into detail when you learned about the theory and implementation of algorithms, but actually deep neural networks have many hyper parameters that need to be decided compared to the past neural networks or other machine learning algorithms. This means we have to go through more trial and error to get high precision. Combinations of parameters that define a structure of neural networks, such as how many hidden layers are to be set or how many units each hidden layer should have, need lots of experiments. Also, the parameters for training and test configurations such as the learning rateneed to be determined. Furthermore, peculiar parameters for each algorithm such as the corruption level in SDA and the size of kernels in CNN need additional trial and error. Thus, the great performance that deep learning provides is supported by steady parameter-tuning. However, people only look at one side of deep learning—that it can get great precision— and they tend to forget the hard process required to reach to the point. Deep learning is not magic.
In addition, deep learning often fails to train and classify data from simple problems. The shape of deep neural networks is so deep and complicated that the weights can't be well optimized. In terms of optimization, data quantities are also important. This means that deep neural networks require a significant amount of time for each training. To sum up, deep learning shows its worth when:
- It solves complicated and hard problems when people have no idea what feature they can be classified as
- There is sufficient training data to properly optimize deep neural networks
Compared to applications that constantly update a model using continuously updated data, once a model is built using a large-scale data set that doesn't change drastically, applications that use the model universally are rather well suited for deep learning.
Therefore, when you look at business fields, you can say that there are more cases where the existing machine learning can get better results than using deep learning. For example, let's assume we would like to recommend appropriate products to users in an EC. In this EC, many users buy a lot of products daily, so purchase data is largely updated daily. In this case, do you use deep learning to get high-precision classification and recommendations to increase the conversion rates of users' purchases using this data? Probably not, because using the existing machine learning algorithms such as Naive Bayes, collaborative filtering, SVM, and so on, we can get sufficient precision from a practical perspective and can update the model and calculate quicker, which is usually more appreciated. This is why deep learning is not applied much in business fields. Of course, getting higher precision is better in any field, but in reality, higher precision and the necessary calculation time are in a trade-off relationship. Although deep learning is significant in the research field, it has many hurdles yet to clear considering practical applications.
Besides, deep learning algorithms are not perfect, and they still need many improvements to their model itself. For example, RNN, as mentioned earlier, can only satisfy either how past information can be reflected to a network or how precision can be obtained, although it's contrived with techniques such as LSTM. Also, deep learning is still far from the true AI, although it's definitely a great technique compared to the past algorithms. Research on algorithms is progressing actively, but in the meantime, we need one more breakthrough to spread out and infiltrate deep learning into broader society. Maybe this is not just the problem of a model. Deep learning is suddenly booming because it is reinforced by huge developments in hardware and software. Deep learning is closely related to development of the surrounding technology.
As mentioned earlier, there are still many hurdles to clear before deep learning can be applied more practically in the real world, but this is not impossible to achieve. It isn't possible to suddenly invent AI to achieve technological singularity, but there are some fields and methods where deep learning can be applied right away. In the next section, we’ll think about what kinds of industries deep learning can be utilized in. Hopefully, it will sew the seeds for new ideas in your business or research fields.
The approaches to maximize deep learning possibilities and abilities
There are several approaches on how we can apply deep learning to various industries. While it is true that an approach could be different depending on the task or purpose, we can briefly categorized the approaches as the following three:
- Field-oriented approach: This utilizes deep learning algorithms or models that are already thoroughly researched and can lead to a great performance
- Breakdown-oriented approach: This replaces problems to be solved that deep learning can apparently be applied with a different problem that deep learning can be well adopted
- Output-oriented approach: This explores new ways on how we express the output with deep learning
These approaches are all explained in detail in the following subsections. Each approach is divided into its suitable industries or not for its use, but any of them could be a big hint for your activities going forward. There are still very few use cases of deep learning and bias against fields of use, but this means there should be many chances to create innovative and new things. Start-ups who utilize deep learning have been emerging recently and some of them have already achieved success to some extent. You can have a significant impact on the world depending on your ideas.
This approach doesn't require new techniques or algorithms. There are obviously fields that are well suited to the current deep learning techniques, and the concept here is to dive into these fields. As explained previously, since deep learning algorithms that have been practically studied and developed are mostly in image recognition and NLP, we'll explore some fields that can work in great harmony with them.
Medical fields should be developed by deep learning. Tumors or cancers are detected on scanned images. This means nothing else but being able to utilize one of the strongest features of deep learning—the technique of image recognition. It is possible to dramatically increase precision using deep learning to help with the early detection of an illness and identifying the kind of illness. Since CNN can be applied to 3D images, 3D scanned images should be able to be analyzed relatively easily. By adopting deep learning more in the current medical field, deep learning should greatly contribute.
We can also say that deep learning can be significantly useful for the future medical field. The medical field has been under strict regulations, however, there is a movement progressing to ease the regulations in some countries, probably because of the recent development of IT and its potential. Therefore, there will be opportunities in business for the medical field and IT to have a synergy effect. For example, if telemedicine is more infiltrated, there is the possibility that diagnosing or identifying a disease can be done not only by a scanned image, but also by an image shown in real time on a display. Also, if electronic charts become widespread, it would be easier to analyze medical data using deep learning. This is because medical records are compatible with deep learning as they are a dataset of texts and images. Then, the symptom of unknown disease can be found.
We can say that surroundings off running cars are image sequences and texts. Other cars and views are images and a road sign has texts. This means we can also utilize deep learning techniques here, and it is possible to reduce the risk of accidents by improving driving assistance functions. It can be said that the ultimate type of driving assistance is self-driving cars, which is being tackled mainly by Google and Tesla. An example that is both famous and fascinating was when George Hotz, the first person to hack the iPhone, built a self-driving car in his garage. The appearance of the car was introduced in an article by Bloomberg Business (http://www.bloomberg.com/features/2015-george-hotz-self-driving-car/), and the following image was included in the article:
Self-driving cars have been already tested in the U.S., but since other countries have different traffic rules and road conditions, this idea requires further studying and development before self-driving cars are commonly used worldwide. The key to success in this field is in learning and recognizing surrounding cars, people, views, and traffic signs, and properly judging how to process them.
In the meantime, we don't have to just focus on utilizing deep learning techniques for the actual body of a car. Let's assume we could develop a smartphone app that has the same function as we just described, that is, recognizing and classifying surrounding images and text. Then, if you just set up the smartphone in your car, you could utilize it as a car-navigation app. In addition, for example, it could be used as a navigation app for blind people, providing them with good reliable directions.
Advert (ad) technologies could expand their coverage with deep learning. When we say ad technologies, this currently means recommendation or ad networks that optimize ad banners or products to be shown. On the other hand, when we say advertising, this doesn't only mean banners or ad networks. There are various kinds of ads in the world depending on the type of media such as TV ads, radio ads, newspaper ads, posters, flyers, and so on. We have also digital ad campaigns with YouTube, Vine, Facebook, Twitter, Snapchat, and so on. Advertising itself has changed its definition and content, but all ads have one thing in common, they consist of images and/or language. This means they are fields that deep learning is good at. Until now, we could only use user-behavior-based indicators, such as page view (PV), click through rate (CTR), and conversion rate (CVR), to estimate the effect of an ad, but if we apply deep learning technologies, we might be able to analyze the actual content of an ad and autogenerate ads going forward. Especially since movies and videos can only be analyzed as a result of image recognition and NLP, video recognition, not image recognition, will gather momentum besides ad technologies.
Profession or practice
Professions such as doctor, lawyer, patent attorney, and accountant are considered to be roles that deep learning can replace. For example, if NLP's precision and accuracy gets higher, any perusal that requires expertise can be left to a machine. As a machine can cover these time-consuming reading tasks, people can focus more on high-value tasks. In addition, if a machine classifies past judicial cases or medical cases on what disease caused what symptoms and so on, we would be able to build an app like Apple’s Siri that answers simple questions that usually require professional knowledge. Then, the machine could handle these professional cases to some extent if a doctor or a lawyer is too busy to help in a timely manner.
It's often said that AI takes away human’s jobs, but personally, this seems incorrect. Rather, a machine takes away menial work, which should support humans. A software engineer who works on AI programming can described as having a professional job, but this work will also be changed in the future. For example, think about a car-related job, where the current work is building standard automobiles, but in the future engineers will be in a position just like pit crews for Formula 1 cars.
Deep learning can certainly contribute to sports as well. In the study field known as sports science, it has become increasingly important to analyze and examine data from sports. As an example, you may know the book or movie Moneyball. In this film, they hugely increased the win percentage of the team by adopting a regression model in baseball. Watching sports itself is very exciting, but on the other hand, sport can be seen as a chunk of image sequences and number data. Since deep learning is good at identifying features that humans can't find, it will become easier to find out why certain players get good scores while others don't.
These fields we have mentioned are only a small part of the many fields where deep learning is capable of significantly contributing to development. We have looked into these fields from the perspective of whether a field has images or text, but of course deep learning should also show great performance for simple analysis with general number data. It should be possible to apply deep learning to various other fields such as bioinformatics, finance, agriculture, chemistry, astronomy, economy, and more.
This approach might be similar to the approach considered in traditional machine learning algorithms. We already talked about how feature engineering is the key to improving precision in machine learning. Now we can say that this feature engineering can be divided into the following two parts:
- Engineering under the constraints of a machine learning model. The typical case is to make inputs discrete or continuous.
- Feature engineering to increase precision by machine learning. This tends to rely on the sense of a researcher.
In a narrower meaning, feature engineering is considered as the second one, and this is the part that deep learning doesn't have to focus on, whereas the first one is definitely the important part even for deep learning. For example, it's difficult to predict stock prices using deep learning. Stock prices are volatile and it’s difficult to define inputs. Besides, how to apply an output value is also a difficult problem. Enabling deep learning to handle these inputs and outputs is also said to be feature engineering in the wider sense. If there is no limitation to the value of original data and/or data you would like to predict, it’s difficult to insert these datasets into machine learning and deep learning algorithms, including neural networks.
However, we can take a certain approach and apply a model to these previous problems by breaking down the inputs and/or outputs. In terms of NLP as explained earlier, you might have thought, for example, that it would be impossible to put numberless words into features in the first place, but as you already know, we can train feed-forward neural networks with words by representing them with sparse vectors and combining N-grams into them. Of course, we can not only use neural networks, but also other machine learning algorithms such as SVM here. Thus, we can cultivate a new field where deep learning hasn't been applied by engineering to fit features well into deep learning models. In the meantime, when we focus on NLP, we can see that RNN and LSTM were developed to properly resolve the difficulties or tasks encountered in NLP. This can be considered as the opposite approach to feature engineering because in this case, the problem is solved by breaking down a model to fit into features.
Then, how do we do engineering for stock prediction as we just mentioned? It's actually not difficult to think of inputs, that is, features. For example, if you predict stock prices daily, it’s hard to calculate if you use daily stock prices as features, but if you use a rate of price change between a day and the day before, then it should be much easier to process as the price stays within a certain range and the gradients won't explode easily. Meanwhile, what is difficult is how to deal with outputs. Stock prices are of course continuous values, hence outputs can be various values. This means that in neural network model where the number of units in the output layer is fixed, they can't handle this problem. What should we do here—should we give up?! No, wait a minute. Unfortunately, we can't predict a stock price itself, but there is an alternative prediction method.
Here, the problem is that we can classify stock prices to be predicted into infinite patterns. Then, can we make them into limited patterns? Yes, we can. Let's forcibly make them. Think about the most extreme but easy to understand case: predicting whether a tomorrow's stock price, strictly speaking a close price, is up or down using the data from the stock price up to today. For this case, we can show it with a deep learning model as follows:
In the preceding image, denotes the open price of a day, ; denotes the close price, is the high price, and is the actual price. The features used here are mere examples, and need to be fine-tuned when applied to real applications. The point here is that replacing the original task with this type of problem enables deep neural networks to theoretically classify data. Furthermore, if you classify the data by how much it will go up or down, you could make more detailed predictions. For example, you could classify data as shown in the following table:
Up more than 3 percent from the closing price
Up more than 1~3 percent from the closing price
Up more than 0~1 percent from the closing price
Down more than 0~-1 percent from the closing price
Down more than -1~-3 percent from the closing price
Down more than -3 percent from the closing price
Whether the prediction actually works, in other words whether the classification works, is unknown until we examine it, but the fluctuation of stock prices can be predicted in a quite narrow range by dividing the outputs into multiple classes. Once we can adopt the task into neural networks, then what we should do is just examine which model gets better results. In this example, we may apply RNN because the stock price is time sequential data. If we look at charts showing the price as image data, we can also use CNN to predict the future price.
So now we've thought about the approach by referring to examples, but to sum up in general, we can say that:
- Feature engineering for models: This is designing inputs or adjusting values to fit deep learning models, or enabling classification by setting a limitation for the outputs
- Model engineering for features: This is devising new neural network models or algorithms to solve problems in a focused field
The first one needs ideas for the part of designing inputs and outputs to fit to a model, whereas the second one needs to take a mathematical approach. Feature engineering might be easier to start if you are conscious of making an item prediction-limited.
The previously mentioned two approaches are to increase the percentage of correct answers for a certain field's task or problem using deep learning. Of course, it is essential and the part where deep learning proves its worth, however, increasing precision to the ultimate level may not be the only way of utilizing deep learning. Another approach is to devise the outputs using deep learning by slightly changing the point of view. Let's see what this means.
Deep learning is applauded as an innovative approach among researchers and technical experts of AI, but the world in general doesn't know much about its greatness yet. Rather, they pay attention to what a machine can't do. For example, people don't really focus on the image recognition capabilities of MNIST using CNN, which generates a lower error rate than humans, but they criticize that a machine can't recognize images perfectly. This is probably because people expect a lot when they hear and imagine AI. We might need to change this mindset. Let's consider DORAEMON, a Japanese national cartoon character who is also famous worldwide—a robot who has high intelligence and AI, but often makes silly mistakes. Do we criticize him? No, we just laugh it off or take it as a joke and don’t get serious. Also, think about DUMMY / DUM-E, the robot arm in the movie Iron Man. It has AI as well, but makes silly mistakes. See, they make mistakes but we still like them.
In this way, it might be better to emphasize the point that machines make mistakes. Changing the expression part of a user interface could be the trigger for people to adopt AI rather than just studying an algorithm the most. Who knows? It’s highly likely that you can gain the world’s interest by thinking of ideas in creative fields, not from the perspective of precision. Deep Dream by Google is one good example. We can do more exciting things when art or design and deep learning collaborate.
And …congratulations! You’ve just accomplished the learning part of deep learning with Java. Although there are still some models that have not been mentioned, you can be sure there will be no problem in acquiring and utilizing them.
Resources for Article:
- Setup Routine for an Enterprise Spring Application[article]
- Support for Developers of Spring Web Flow 2[article]
- Using Spring JMX within Java Applications[article]