Neuro-Symbolic AI – Unlocking the Next Phase of AI

Don't let the AI hype fool you. Six years after Elon Musk warned AI-researchers were "summoning the demon," the field is still decades away from achieving true general AI that's autonomous and cross domain. In fact, as Oren Etzioni, CEO of the Allen Institute for AI, notes, "it turns out that we don’t even know, in some sense, the right questions to ask” to get us there.

The reality is today's AI is limited to narrow tasks within single domains where there's a large volume of annotated training data and experts to train models. When this is the case, as it is with general image classification thanks to ImageNet, AI techniques are highly accurate at taking an input and predicting an output. For example, if you provide a photo as input, an AI will tag the contents of the photo as output. Or if a customer writes a support question on a website as input, an AI will classify the customer's intent and trigger a specific dialogue flow based on it.

I don't want to downplay the significance of low-cost, automated predictions and classification, it's just not the "intelligence" most people assume AI is capable of. Neural-based techniques still can't reason nor can they infer causal relationships to explain why something happened or what might happen next, which are hallmarks of even low level intelligence.  

This "intellectual" immaturity is a reason why AI is everywhere but its impact is almost imperceptible to most of us. Music playlists, driving directions, search results, and language translation, are all examples of services already made a bit better by AI.  

Over the next five years, AI will continue to scale and improve even more services, remove more tedium from your work and personal life, and usher in more personalized experiences. Every part of the economy will get an "AI boost" just using what's possible today. Electronics companies such as Sony will enable novice photographers to shoot like pros with AI enabled image sensors that adjust settings based on the subject(s) actually in frame. Retail organizations will automatically tailor app designs specifically to your individual preferences. And Robotic Process Automation and Global Systems Integration firms will optimize business processes by automating out tedious tasks such as data entry, thus leaving you with higher value judgement and creativity related work.

This is all fantastic, but to unlock higher level "intelligence," AI needs to work on less data, generalize outside its training set / domain, and be more explainable (i.e. explain how an output was determined).  

"Deep learning techniques thus far have proven to be data hungry, shallow, brittle, and limited in their ability to generalize." - Gary Marcus, a leading AI scientist

Researchers at the MIT-IBM Watson AI Lab believe Neuro-Symbolic AI is potentially a way to address these neural-based AI limitations. Their solution is to mash-up two AI techniques (i.e. Neural and Symbolic) to achieve a 1+1=3 effect . Used together, neural and symbolic techniques can reason as to why something happened or what might happen next based on the properties of what the AI is analyzing (e.g. a ball is round and round objects roll when on uneven surfaces). If that's not impressive enough, it does this using less data and in a more explainable and generalizable way!

In late 2019, a collaboration between MIT CSAIL, IBM Research, Harvard University, and Google DeepMind, published a paper (CLEVRER: Collision Events for Video Representation and Reasoning) describing how Neuro-Symbolic AI was used to "recognize objects in a video, analyze their movements, and reason about their behaviors".

This achievement is significant because it proves computers can build cognitive models similar to how humans do and answer questions that are:

• Descriptive: e.g. “What is the material of the last object to collide with the cyan cylinder?”
• Explanatory: e.g. “What is responsible for the collision between rubber and metal cylinders?”
• Predictive: e.g. “What will happen next?”
• Counterfactual: e.g. “What will happen without the cyan cylinder?”

While Neuro-Symbolic AI is just emerging from research and will evolve over many years, just as today's narrow AI has, this research proves the next step towards general AI can be taken and broadens the solutions AI can be applied to in the coming years. Neuro-Symbolic AI almost certainly isn't the panacea that fixes all neural-based ills. So time will tell just how big a step foward we can take with it.