The Scope and Challenges for Deep Learning

I’ve been using machine-learning methods for more than two decades for a variety of problems, the most challenging of which has been to design a set of algorithms that learn from data how to trade securities profitably on their own and with minimal human input. I started this research to answer a basic question, namely, whether machine-learning algorithms could be shaped to make investment decisions comparable to the best humans with access to the same data. I hitched myself to the machine-learning post out of the conviction that financial markets are too complex, subtle, and ‘‘noisy’’ to be described by humans in simple terms, and that they evolve over time along multiple dimensions. By noisy, I mean that seemingly identical market conditions at different times can result in very different outcomes. My thinking was that machine-learning algorithms, by virtue of their ability, in principle, to learn or generalize from large amounts of noisy data, should have advantages over humans when it comes to making decisions more frequently. The broad approach I took is commonly referred to as ‘‘machine learning’’ (ML). The creative work of crafting useful descriptions about the problem that are fed to the machine is done by humans, based on intuition and prior theory if it exists. The computer then learns to generalize from this type of curated input. For example, I might take a few price series and do some calculations on them that constitute the curated input fed to the machine. If it turns out that a calculation such as ‘‘increasing volatility’’ of the series (which may be indicative of a jumpy market) is generally bearish for stocks, the ML algorithms should be able to infer this by doing some arithmetic on the subsequent outcomes associated with the market under jumpy conditions. But what if instead we gave the computer just the raw data—prices, news, earnings, and macro numbers—and asked it to learn from them directly, without any human curation of the raw data? Sound like science fiction? Perhaps, but before we dismiss it as outlandish, it is worth considering the surge in interest among the Internet giants such as Google and Facebook, which is matched in fervor, if not financial resources, by smaller innovators in the area of ‘‘deep learning.’’ The UKbased company DeepMind was acquired by Google for roughly $400 million last year. What was so significant about this acquisition? Consider the following experiment on learning to play video games. With access only to the video output on the screen, game controllers, and the game score—the only feedback available for learning—a deep learning based program learned how to play three games better than humans. It is hard to understate the significance of this feat. The program required no prior knowledge about the problem it would be trying to solve and no preprocessing of inputs by humans. The machine ‘‘saw’’ the same input that a human would see, in the same form—sequences of pixels—and learned to play the games very well. What is the essential difference between traditional ML and deep learning that enables this ‘‘different in kind’’ capability? And perhaps more tantalizingly, what are the promises and challenges associated with this difference in kind? A fundamental difference is in how the problem is formulated and how the ‘‘decision surface’’ is generated from the data. Unlike standard ML approaches (such as the ones I’ve used to develop financial prediction systems) in which the raw data are curated into features that direct and constrain the algorithms’ explorations in a specified manner, there is zero to minimal pre-processing of the raw data in the deep-learning context. Raw data on, say, pixels of the screen in a video game or a digitized facial image or raw text are used to learn useful features implicitly. Raw data and nothing else. Features derived from pixels might be lines, shadows, distance between similar