Guest Opinion: Time to go Deep on Deep Learning

  Ravi Sawhney,   Global Head of Credit FI Electronic Trading

Ravi Sawhney, Global Head of Credit FI Electronic Trading

Time to go Deep on Deep Learning

Artificial Intelligence is no longer something to be associated with the future. It is very much in the here and now, and is being employed by corporations and governments throughout the world to make decisions that impact many aspects of our lives. Mainstream attention has exploded in recent years on the back of a rise in Machine Learning and Deep Learning applications and technologies. However, increased attention doesn’t necessarily correlate to understanding. As the lines continue to blur between decisions made by machine vs. human, it will become critical for us to comprehend the inner workings of AI. Without this understanding, how can we know who is really in control?

Machine Learning In Practice

 The most common form of Machine Learning used in applications today is supervised learning. This form of learning requires a vast amount of labelled data whereby a real-world truth has already been established. This data is then used to train a machine learning algorithm which can then be applied to a situation (never observed before) to predict a so-called “outcome classifier.” Questions that Machine Learning programs could address might include: Is this a picture of a dog? What might the next word be in this sentence? Should I (the software running on a self-driving car) apply the brakes? Is this credit card transaction fraudulent?

 How Machine Learning Differs From Deep Learning

 A core difference between Machine Learning and Deep Learning is in the feature selection process, which is the function by which data is chosen in creating a predictive model. In Machine Learning, domain expertise is required to code the inputs used to build a model. For example, let's say you are building a model for facial recognition. You might start by determining where the eyes, nose and mouth are located. Doing this is something we humans can do very easily; however, for a machine, it’s not so simple. After a domain expert selects the relevant features, those features can then be expressed numerically for each image in the training set. This data can then be fed into a machine learning algorithm to create a model which can make decisions on an image it has never encountered before.  

If a Deep Learning approach were applied to the same facial recognition challenge, a developer would simply load the image pixels, represented numerically, directly to the model itself. No prior domain experience of facial recognition would be required and the program  would not need to be pre-loaded with data on eye, nose or face measurements. Instead, a type of Artificial Neural Network (ANN), inspired by the workings of the human brain, would be used to automatically find the right features from the data to create the model.

Given that domain expertise is not a requirement to create state-of-the-art models, people are (understandably) becoming both very excited and nervous as Deep Learning methodologies improve. But whether we like it or not, the reality is that in five years or less, advances in Deep Learning will continue to move forward at a blistering pace (For more on why I am so confident in my prediction, see my previous post
here).

Deep Learning Applied to Financial Services

Deep Learning techniques are being actively explored today, if not beginning to be deployed, in Financial Services. One of the use cases gaining considerable attention is the deployment of Deep Learning in quantivative investment strategies.To understand the potential, think back to the facial recognition example as applied to the challenges faced by portfolio managers seeking to outperform a benchmark. By using Machine Learning, the investment manager would have to choose which variables (i.e., the features) were most relevant in driving stock price performance. Doing so requires a huge degree of skill and domain expertise and, of course, comes with a heavy likelihood of failure.

Now imagine a Deep Learning model used to tackle the same challenge. The model could take all possible data sets available and ‘figure out’ which ones are more relevant in generating alpha. Furthermore, in addition to using structured data (e.g. ratios from a company's balance sheet, income statements, profit metrics, etc.) the model could also absorb vast amounts of unstructured data such as transcripts from quarterly investor calls.

What Could Go Wrong?

Deep Learning systems, like their human creators, will inevitably make mistakes. These setbacks will likely get more serious as applications extend far beyond mistagging someone in a Facebook photo. Imagine when they cause a car accident, misdiagnose a condition based on a faulty read of an MRI or trigger significant financial losses through a bad trade. When this happens, we will want to understand the cause of the mistake and what improvements should be made to prevent a similar occurrence. But doing that forensic work represents the core challenge of using Deep Learning models. Quite simply, today, we just do not know how to fully interpret the decisions made by Deep Learning systems.

Fortunately, gifted technologists in both the private and public sectors are stepping up to address the challenges imposed by Deep Learning systems. Financial services start-ups such as LogicalGlue, which seeks to combine Deep Learning techniques with more transparent algorithms, have made progress in demystifying AI-based programs. Moreover, the US Department of Defense’s DARPA has launched a program to better understand an emerging generation of “artificially intelligent machine partners.”

 Clearly, more needs to be done. While we all see the potential of computer assisted intelligence in financial, health and many other applications, we need to have greater understanding in how computers reason. Otherwise, our ability to interpret Deep Learning outcomes will diminish over time.

 
Ravi Sawhney
Global Head of Credit FI Electronic Trading
Bloomberg LP
(All views in this article are my own)