Artificial reasoning going beyond machine learning to achieve greater intelligence

The success of deep learning and other machine learning (ML) algorithms in real world applications has been a major boost for the artificial intelligence (AI) research community. These applications can be classed as narrow AI, and the long-term goal is to achieve general AI, or artificial general intelligence, where the AI machine can replicate the intelligence of a human being. There is a recognition that to go beyond current deep learning neural networks we need to go beyond backpropagation: this is the 1985-86 breakthrough technique which applies calculus and the chain rule to make corrections to synaptic strengths that lie in deep (or hidden) network layers. The AI research community is actively engaged in exploring mechanisms in ML that correspond more closely to neuroscience and our understanding of real brain activity, see for example the 2020 NeurIPS workshop “Beyond Backpropagation”.

As has been pointed out by many observers and practitioners, AI today is not very intelligent. There is an enormous gap between narrow AI and general AI, and it is surely helpful to have an intermediate next step to focus on. Achieving an artificial form of human reasoning looks like a worthwhile candidate. A none too technical definition of reasoning is “the action of thinking about something in a logical, rational way” and drawing of inferences or conclusions through reasoning. Where today you may need to train a deep neural network with millions of images of cats to recognize a test cat, a reasoning AI may see one or two images of a cat and reason that a third test cat is also a cat.

I’ve noted references to “machine reasoning” in the past few years. The earliest I’ve seen is a paper by Leon Bottou in 2011 “From Machine Learning to Machine Reasoning” and available on arXiv. Also on arXiv is a 2020 paper “Machine Reasoning Explainability” from Ericsson Research by K Cyras et al. The idea of machine reasoning is to augment ML with additional systems: knowledge banks, symbolic reasoning, expert systems, heuristic systems, probabilistic systems, Bayesian systems etc., and achieve a pragmatic approach to take machine learning to the next step. 

The telco industry is currently heavily engaged in researching intent-based network management (hence Ericsson’s interest), where you tell the system what needs doing (the goals or intents) and it has sufficient intelligent automation built-in to figure out the optimal way to achieve the task. As 5G+ networks bring in greater complexity, this AI based automation is seen as a necessary step for the industry to increase its efficiency. 

Currently there are few AI research papers that have machine reasoning in their title. Appearing in arXiv I counted seven in the decade 2011-2021 (in contrast there were 6,928 ML titled papers); four of the machine reasoning papers appeared in 2020, and there is one paper with “artificial reasoning” in the title since arXiv began in 1991, quite an intriguing one by Mark Burgess.

Tied to the quest to create artificial reasoning is solving one of mankind’s biggest scientific challenges: what is the mechanism in the human brain that enables thinking? How does the human brain reason?