Brainpower: Man vs. Machine by Benjamin Ruegsegger, AllianceBernstein
Throughout history, humans have been the most intelligent beings on earth. But is this about to change? The advancement of neural networks could be the single most important development in helping machines think more like humans. Investors should take note.
The human body is incredibly adept at sensing the world around it—largely thanks to a complex nervous system manned by a massive network of neurons. These neurons “fire” when they’re stimulated by inputs such as images or temperature. They act independently, but the network processes information collectively and efficiently. The human brain is the most complex neural network, with an estimated 80–100 billion neurons, each with 1,000 connections.
Building a Brain: Helping Machines Learn
Qualivian Investment Partners Up 30% YTD; Long ORLY Thesis
Qualivian Investment Partners commentary for the second quarter ended July 30, 2020. Q2 2020 hedge fund letters, conferences and more “Short-term investors will accept a 20% gain because they didn’t spend the time to develop the conviction and foresight to see the next 500%.” - Ian Cassell Executive Summary Readers of investment letters fall into Read More
It’s not easy to mimic the human brain. Stanford University professor Andrew Ng took an early stab at it in 2011 with Google’s Deep Learning project—later called “Google Brain.” The initial setup had 1,000 computer servers and was roughly equivalent (as measured by the number of connections) to a honeybee’s brain. The cost? A cool $5 million. After Google Brain was fed YouTube images for three straight days, it could identify the faces of a human and a cat. Recently, another tech firm, Nvidia, announced hardware that had similar capabilities but cost only $33,000, bringing brainlike architecture (or thinking) to the masses.
Google Brain is an example of machine learning—using hardware and software to solve problems through “learning” instead of through rule-based instructions. Artificial neural networks (ANNs), modeled after the complex neural connections in the human brain, have been one of the most successful of these approaches so far.
Early Applications of Neural Networks
Neural networks are already at work in more places than you might think. Facebook Inc (NASDAQ:FB) and Google Inc (NASDAQ:GOOGL) (NASDAQ:GOOGL) use them in image searches and to dynamically target advertisements to users. Car manufacturers use them to process images from onboard cameras feeding safety features like lane-drift warnings and pedestrian detection. In fact, any computer application that uses pattern recognition or image analysis is perfect for a neural network framework. And there’s more and more data to work with: the amount of analyzable data in the world is growing exponentially (Display).
Most efforts so far to imitate the human brain have centered around software, but in recent years researchers have tried their hands at hardware too. IBM is working on a computer chip that has better sensory capabilities and uses less power than traditional chips. The company’s long-term goal is to create a system of 10 billion neurons that consumes 1 kilowatt of power and has a volume of less than two liters. International Business Machines Corp. (NYSE:IBM) has committed $3 billion over the next five years to semiconductor research in areas such as new chip architectures. Other companies such as QUALCOMM, Inc. (NASDAQ:QCOM) are making similar investments.
Who’s Capitalizing on Machine-Based Learning?
Google and Facebook have been noted for their efforts to recruit thought leaders applying this technology. It’s not surprising that these tech-savvy companies are leading in machine learning. But the applications are likely to span multiple industries.
One notable example is the race to design the most effective semi- and fully autonomous driving systems. Given the massive volume of images that must be captured and processed for these systems, the hardware vendors, auto parts suppliers and original equipment manufacturers that can design the most efficient and accurate systems will see quicker times to market and fewer incidents. Neural network frameworks may be a real differentiator.
Smaller, entrepreneurial finance companies are also using neural networks to create better lending models by more accurately predicting credit behaviors, such as defaults. Fraud detection—where pattern recognition is particularly useful—is another emerging application.
This technological shift is changing how investors gauge a company’s potential in terms of its intellectual property. Traditionally, one measure of this has been a company’s patent portfolio. Today, it may be better to know how many machine-learning experts are on staff. More talent usually leads to better products and smarter capital spending—driven by neural networks’ greater efficiency versus traditional data analysis. This is especially true for companies whose products involve image and pattern recognition.
Even today, our understanding of how the human brain works is still limited. The more we understand, the more complex we can make the systems that mimic the brain. True artificial intelligence may still be a few years away, but progress is being made on this incredible journey. Today’s machines are more adaptive than ever—mainly because of neural networks. As company assets, experts in machine learning are quickly becoming as important as patent portfolios.
The views expressed herein do not constitute research, investment advice or trade recommendations and do not necessarily represent the views of all AllianceBernstein portfolio-management teams.
Benjamin Ruegsegger is Portfolio Manager—Growth Equities at AllianceBernstein Holding LP (NYSE:AB)