Accelerating the Power of AI with Neural Networks

Bob Friday, CTO and Co-founder, Mist
Originally posted on aitrends

Artificial Intelligence, Machine Learning and Neural Networks Defined

Using the Turing Test as a qualifier, Artificial Intelligence (AI) is defined as a software solution that performs a task on par with a human domain expert. When IBM’s Watson system played Jeopardy with former Jeopardy champions, much of the world saw the first real example of AI. Now, deep learning is enabling solutions that can interpret MRI images on par with doctors and operate buses on par with human drivers (e.g. Las Vegas Self Driving Shuttle).

Machine Learning (ML) is the basic foundation of AI comprised of the algorithms and data sets used to build an AI solution. In order to create a true AI system that can pass the Turing Test, the ML subset must be constantly improving with new sets of data and ongoing developments to the algorithms. While there are many different algorithms that have been in the ML toolbox for decades, it is only recently (circa. 2014) that the deep learning and neural network algorithms have taken a significant leap forward in performance due to the availability of large-labeled data sets for training and low-cost compute and storage.

Neural Networks are Accelerating Machine Learning

Thanks to the fast improvement of computation, storage and distributed computing infrastructure, ML has been evolving into more complex structured models like Deep Learning (DL), Generative Adversarial Network (GAN) and Reinforcement Learning (RL) – all using neural networks. Supervised neural networks are algorithms that can differentiate and make judgements based on image or pattern recognition, after being trained with labeled data. The concept of neural networks has been around for more than forty years, however, it was near 2014 that deep learning and neural networks began to disrupt different segments and bring us closer to passing the Turing Test. Thanks to today’s data gathering capabilities, and sheer volume of said data, neural networking is one of the driving trends in successful ML execution.

Deep learning refers to a set of artificial neural network-based ML models that mimic the working mechanisms of neurons and the nerve network of the human brain. There are two kinds of popular neural network models: the Convolutional Neural Network (CNN) model, which is widely used in different image related applications like autonomous driving, robot, image search, etc., and the Recurrent Neural Network (RNN) model, which is empowering most of the Natural Language Processing-based (NLP) text or voice applications, such as chatbots, virtual home and office assistants and simultaneous interpreters.

Generative Adversarial Network (GAN) is a type of ML technique composed of two deep neural networks competing with each other in a zero-sum game framework. GAN runs typically in the unsupervised fashion; thus, it can help reduce the dependency of deep learning models on the amount of labeled training data.

NLP is another algorithmic trend that is driving ML advancement, particularly in the area of virtual home and office assistants. Similarly to neural networks, NLP is algorithmic based vocal- and word-based recognition. As more AI companies adopt these trends and execute on top of their ML foundation, they will be successful.

Key Considerations in Building an AI System

A solid data pipeline and a great data science toolbox are key to building an effective AI-driven system. We’ve only recently gained access to nearly unlimited compute power and storage in the cloud, which has, in turn, allowed for incredible data collection and analysis. With the right volume and quality of data, as well as the nurturing of data science programs, ML will advance quickly and bring companies closer to achieving true AI.

Almost any college graduate can build and train a deep learning model using tools such Python, TensorFlow and Keras. To bring an AI solution to production, you need tools such as Spark, Kubernetes and Docker to allow the collection and creation of large labeled datasets and data pipelines

There are many open source tools, like TensorFlow, Keras, and Mllib, which dramatically reduce the effort and knowledge required of building a ML – even DL – model, but bringing a solution to production requires the whole ecosystem of AI primitives, including data acquisition and labelling, data processing pipeline, model execution, post deployment validation and continuous model improvement.

In addition, there are other factors determining the success of an AI solution. These include how to leverage and integrate human knowledge and heuristics while developing machine intelligence; how to build human trust in the step-by-step process of automation, augmentation and autonomy; and how to accelerate knowledge learning and sharing across different customers without compromising individuals’ privacy information.

Real-World Use Case for Neural Networks and ML

As virtual home assistants have become commonplace across America, virtual assistants for the enterprise are gaining traction as a prime use case for the predictive capabilities of neural networks and ML algorithms. For example, Dartmouth College, an Ivy League university, has implemented an AI-driven virtual network assistant, Marvis, which provides insight into Wireless LAN (WLAN) behavior and offers expert guidance for rapid Wi-Fi troubleshooting.

Marvis uses NLP to provide Dartmouth’s network administrators with answers to questions such as, “How are the Wi-Fi access points in Baker-Berry Library performing?” With each question posed, the assistant leverages its neural network and becomes more accurate and more confident over time, improving the ML data set and resulting insights. Dartmouth is also leveraging an AI-driven RF planning system that automatically learns and optimizes the Wi-Fi channel and power settings by leveraging reinforcement learning with the reward being improved user experience on the network that drives the learning algorithms.

We are seeing the convergence of several different technologies such as compute, storage and large data sets that are enabling AI, disrupting segments in our society involving images, voice, healthcare and automotive with real world implementations. As adoption continues, and AI becomes more advanced, we will see further developments in AI that, ultimately, disrupt our daily lives.