Most people familiar with the AI of today will think of deep learning as the canonical approach to building intelligent systems. This data-hungry approach has taken over the industry, and achieved stunning successes, especially in perceptual tasks like computer vision and speech recognition.
Deep learning uses multi-layered neural networks, which are really just a fancy way of generating mathematical functions that map input data onto outputs by detecting probabilistic patterns in training datasets. They do not require any upfront specification of exactly what they are analyzing; they generate mathematical functions without using any symbolic representation of the world. Deep learning and similar techniques are sometimes called nonsymbolic AI.
Before deep learning took over, there was an earlier version of AI, symbolic AI, which was the dominant approach in the field from the 1950s to the late 1980s. This approach has fallen out of favor, partly because it failed to live up to its hype, and also, I would argue, because many of the efforts tried to use symbolic AI to achieve artificial general intelligence (AGI), AI that could do anything a human intelligence could (CYC for example). Had symbolic approaches been developed in more constrained settings, perhaps they would be in broader use today.
Nonsymbolic approaches such as deep learning and its simpler less data-hungry siblings such as logistic regression aren’t always appropriate for development of artificial intelligence, especially in business settings. Deep learning requires datasets that are far larger than any we will see in the space we’re working in. And they don’t easily support the kind of if-then modeling required by recommendation engines. The result of unthinkingly applying the dominant nonsymbolic approach has limited AI’s impact on business. Look around at how it’s being used today and you’ll see a wealth of predictive models that don’t offer all that much practical guidance, because they haven’t been wrapped up in domain-aware capabilities useful to humans.
I believe, as do some others working in AI today, that to produce useful artificial intelligence we require a hybrid approach that combines symbolic and nonsymbolic methods. Constraining one’s self to nonsymbolic supervised machine learning techniques just doesn’t provide short-term business value, especially as few businesses have the data volumes necessary to train deep learning models. Instead, pragmatic AI developers will find that developing symbolic representations of their business domains alongside a targeted use of relatively simple machine learning algorithms will accelerate their path towards useful AI.
For development of the Incantata coaching system, we are using a variety of nonsymbolic machine learning algorithms alongside development of an ontology specific to the coaching and counseling domain, demonstrating that symbolic and nonsymbolic methods can co-exist. This approach could be used in a variety of domains, and often implicitly is, as many business AI teams find themselves combining heuristic or machine-generated rules operating on domain ontology elements with machine learning algorithms in order to produce useful intelligence. A broader recognition of the techniques and possibilities of hybrid AI is needed so that data science teams and those who use their work can break out of the fixation on deep learning.