How Artificial Neural Network Types Can Change Business

Imagine a business tool that has the know-how of the human brain and can boost personalization, cut time on menial tasks and improve service delivery at large. Think it’s just science fiction? Think again. Artificial neural networks — artificial intelligence networks modeled after the human brain — are gearing up to have a huge impact on businesses everywhere. In fact, many companies, whether they know it or not, are already are taking advantage of neural networking technology.

Those in the know may be directly taking advantage of ANN technology such as Google Cloud Speech-to-Text, a tool released this year that helps developers convert audio to text by applying neural network models using an application programming interface. The tool recognizes 120 languages and can be employed for purposes such as transcribing audio from call centers.

Microsoft also has some skin in the game. The company has already explored imaging deep neural networks to classify malware and has recently previewed Project Brainwave, a hardware architecture that makes real-time AI calculations. The tool is already being tested by at least one electronics manufacturing solutions provider for an automated optical inspection system to scan products on the assembly line for defects at blazing speed. Meanwhile, IBM has released a beta of its Neural Network Modeler in Watson Studio that lets developers use a visual approach to designing deep learning architectures for image, text and audio data.

The truth is, however, that plenty of companies that don’t know that some key business processes are backed by ANN technology. Marketers aren’t necessarily aware that Facebook’s DeepText — a deep-learning-based text understanding engine that leverages neural network architectures — is used to better target the ads they place so that they appear to users who are most likely to find them relevant.

DOWNLOAD: Learn more about how hybrid IT can help your business adopt emerging technologies!

What Are Artificial Neural Networks?

ANNs are one subset of artificial intelligence, “a broad term that encompasses all the technologies or techniques that enable computers to mimic the cognitive functions of humans, such as planning, vision, understanding language, recognizing sound, speech, learning and problem-solving,” says PR Krishnan, executive vice president and global head of enterprise intelligence automation at Tata Consultancy Services.

Artificial neural networks created from artificial neurons are the building blocks of deep learning, Krishnan says. Deep learning is one of the techniques classified under the banner of machine learning, which predominantly involves building a model from existing training data and applying that model to new data for prediction or estimation. With deep learning, there are multiple layers of neural networks, with each one learning some aspect of the overall problem.

Put simply, neural networks are loosely modeled after the human brain. In humans, neurons act as information highways between different areas of the brain. In artificial neural networks, neurons receive, process and transmit information, too.

“Deep learning is extensively used in highly complex applications such as computer vision, speech translation and similar skillsets,” says Krishnan. Shallow learning also is a machine learning technique to learn human behavior over time for purposes like product recommendations on many e-commerce websites.

The Types of Neural Networks and When to Use Them

Neural networks come in a variety of types that can be applied to separate use cases:

• Convolutional neural networks: Similar to ordinary neural networks, CNNs differ in that they “make the explicit assumption that the inputs are images,” according to GitHub. This makes these most useful to analyze and classify visual imagery.
• Recurrent neural networks: Real human brains store memories and use them as a starting point to interpret new information. In a way, RNNs do the same thing. RNNs are networks with loops that allow information to “persist,” and are therefore useful in recognizing speech.
• Gated recurrent unit: Similar to short-term memory, GRUs add a gating mechanism to RNNs. They can be trained to retain information and remove what is not relevant to a current prediction. They are most effective in analyzing smaller data sets.
• Long Short-Term Memory: LTSMs are RNNs that are capable of learning long-term dependencies. They are useful for solving a variety of problems, making them very popular.
• Reinforcement Learning: These are “goal-oriented algorithms” that “learn” how a piece of software should best operate in a specific environment by performing actions and examining the outcome, receiving rewards for optimal interactions and vice versa. Over time, these algorithms can uncover an optimal pathway. This is the type of AI that famously beat the world’s best human player in a game of Go.

“The key question [when choosing a type to investigate] is, ‘What business problem are you trying to solve?’” says Randy Hlavac, an early experimenter with neural networks. Hlavac currently works as a lecturer at the Medill School at Northwestern University and is CEO of consulting company Marketing Synergy, which works with companies to apply predictive technologies for business needs.

According to Hlavac, the types of neural networks that will be most frequently used by companies in the future have to not only solve a business problem but also achieve high accuracy and provide solid and measurable ROI.

“The extensive use of CNN in the computer vision space is a good example of this,” says Krishnan. In contrast, reinforced learning, which is used heavily in gaming, isn’t commonly used in business applications because its technology limitations negatively impact ROI, he says.

According to AI startup adviser Steve Ardire, the main neural networks that business users need to know about today are CNN for image recognition and computer vision use cases and RNN, which is designed for natural language processing use cases. Additionally, supervised learning — where the machine is trained using well-labeled data — is primarily used for such deep learning techniques. Ardire notes, however, that “the future is unsupervised learning where machines can infer what they don't know about and are given no positive or negative reinforcement.”

Cloud and Big Data Make Artificial Neural Networks Possible

The main issue with scaling supervised learning techniques is a lack of labelled and tagged training data, but the good news is that there is a lot more data than ever before to apply to training neural networks.

In fact, companies can work with the data kings — Google, Amazon, Facebook and so on — to gain access to anonymized data that fits the profile for business use cases and apply that data in combination with their own cloud-hosted data.

“The Big Data we have now and cloud computing are changing everything,” says Hlavac, noting that the cloud is capable of processing the enormous amounts of data necessary for training these networks.

Ardire agrees: “Machines can attend to vastly more information and more complex processes than human beings,” he says. “Data is the new oil and AI is the refinery.”

VIDEO: See how businesses are gaining cost savings and flexibility in the cloud!

Challenges Persist for Artificial Neural Networks

Despite these advances making ANNs more possible than ever, there are still some issues preventing the widespread use of neural networks, including the higher computational costs that come with processing large loads of data, Krishnan points out. Other challenges include avoiding bias within existing data, and attaining higher accuracy from less training data can also be a concern.

Meanwhile, further concerns loom large over the future of artificial neural networks, including a shortage of available AI talent, a lack of standardization for AI frameworks and industry data silos that could slow growth.

But once industries and companies are able to jump these hurdles, end users can expect a better granular and holistic experience for common services provided today, according to Krishnan. He posits that personalization could reach new heights, including vendors providing hyper-personalized, context-based recommendation engines to replace more traditional recommendations. Chatbots could get a boost as well, transforming from narrow AI-based chat engines that cater to specific tasks to a personalized chat agent that could reflect humanlike emotions and intelligence.

So how do nontechnical companies get started? Some businesses are already experimenting with artificial neural networks by turning to open-source frameworks to build AI-inspired solutions. Such frameworks, according to Krishnan, “have been able to democratize and package the various facets of deep learning into one framework, making it easier to leverage.”