There have been lots of comments about Google’s recent acquisition of artificial intelligence (AI) company DeepMind and what Google had in mind with this acquisition.
This was not Google’s first foray into artificial intelligence. Google’s translation and speech technologies rely on artificial intelligence. For example Google launched the Quantum QTM Artificial Intelligence Lab to “build better models of the world to make more accurate predictions.” As part of the initiative, Google partnered with NASA. The NASA Ames Research Center is hosting the Quantum Artificial Intelligence Lab and is housing a quantum computer from D-Wave Systems and the Universities Space Research Association (USRA).
In conjunction with this acquisition FICO recently added its interpretation in a blog with the headline: Deep Learning Analytics: The Next Breakthrough in Artificial Intelligence?
Here is what FICO had to say about this subject:
With countless analytic advances based on our understanding of how the brain works, many of us in the data science field were intrigued to hear about Google’s recent acquisition of artificial intelligence (AI) company DeepMind. It reinforces the potential of a new field within AI, that of “deep learning.” Indeed, we here at FICO are well aware of the business applications of artificial intelligence, having pioneered the use of these methods to solve business problems, most notably our fraud detection software that utilizes neural network models.
Much like neural network fraud models, deep learning research focuses on mimicking how the brain learns. A neural network model utilizes “shallow” learning where the inputs are complex domain-specific variables. By contrast, deep learning allows simplistic inputs but leverages deep networks to learn complex relationships between these simpler inputs. In this way, deep learning is trying to get closer to the intricacy of the human brain.
Looking at model architecture helps to illustrate the difference. The figure below represents the architecture of a typical neural network. The “hidden layer” is the mathematical core of a neural net. It selects the combinations of inputs (e.g., dollar amount, transaction type) that are most predictive of the output (e.g., likelihood of fraud).
Shallow Learning (Fraud Example)
A single hidden layer is typically what classifies this network as shallow. By contrast, deep learning networks include multiple hidden layers between the input layer and the output layer. This increased complexity means we can solve more multifaceted problems using larger, more complex data sources – a huge plus in these days of Big Data. For those who would like to dig deeper into these differences, I encourage you to read my more in-depth analysis in a recent FICO Labs Blog post.
As an example of complex problem-solving, deep learning shows strong promise in the areas of image and video analytics. Analytic recognition of images is a rather multidimensional process, involving the consideration of multiple features like edges, pixel clusters and shading that would need to be automatically learned.
Of course, deep learning is in its relative infancy when compared to its shallow learning counterparts. At FICO, we’ve honed our fraud models for more than two decades, and in fact, the use of less complex shallow networks provides multiple benefits, including fast computation, and even more importantly, narrower scope to prevent over-training of the models. This results in robust models that can make quick decisions – paramount for fraud detection.
As the potential for deep learning is being recognized, it’s no wonder that Google and others are so interested in these methods to leverage their Big Data stores. FICO continues to research these types of advances to ensure that it is constantly building the most predictive analytics for clients.