Smarter than the Average Robot: Achieving Common Sense in AI

Artificial intelligence has come a long way in recent years. From beating world champions at complex games like chess and Go, to powering helpful voice assistants like Siri and Alexa, to recommending products and content, AI is transforming our world. However, despite all the hype, most AI today lacks what even young children possess – common sense. Teaching machines this basic but important ability remains one of the biggest challenges in the field of AI.

What is Common Sense and Why is it Important?

Common sense is the basic understanding about the everyday world that most people accumulate through their daily experiences and interactions. It encompasses knowledge about:

• Objects – understanding what common objects are, what they are used for, and their typical properties. For example, knowing that a cup can contain liquid without leaking, but a sieve cannot.
• People – understanding human behavior, motivations, social dynamics and unwritten rules that govern everyday interactions.
• Events – understanding the physics of the world, causality between events, likely results or consequences of actions. For example, knowing that dropping a glass will likely cause it to fall and break.
• Space and time – understanding sizes, distances and durations in the physical world. For example, knowing a human cannot walk from New York to Beijing.

This kind of “world knowledge” that humans gather informally comes instinctively to us and forms the basis for understanding language, behaving reasonably, and making sound decisions. Without common sense, AI systems struggle to operate normally in the open world alongside humans.

Lack of common sense makes today’s AI brittle. Advanced systems that can caption images or generate text may hallucinate or output something nonsensical in edge cases it hasn’t seen before. Self-driving cars still make embarrassing mistakes in normal driving situations due to lack of common sense. Robots cannot yet manipulate objects and navigate environments as flexibly as humans and animals.

Endowing machines with common sense reasoning would make them far more capable, useful and trustworthy. It is key to achieving Artificial General Intelligence (AGI) – the holy grail of human-like adaptability in machines. That’s why common sense remains one of the most important frontiers in AI research today.

The Long Road to Common Sense

The quest to formalize and impart common sense to machines has confounded AI researchers for decades. Unlike narrow domains like games, mathematics or molecular interactions, common sense encompasses vast, informal, contextual and unstructured knowledge about the everyday world.

Earlier approaches relied on manually encoding common sense facts and rules into knowledge bases like Cyc. This proved far too small, brittle and labor-intensive to scale. The Winograd Schema Challenge was proposed as an alternative benchmark for common sense reasoning. However, early attempts based on manually crafted rule-based systems failed to make progress.

The rise of big data and deep learning in the 2010s enabled statistical learning approaches. Large corpora like ConceptNet, Atomic, VisualGenome and SWAG compiled commonsense knowledge and reasoning examples at scale. Language models like BERT, GPT-3 and PaLM showed some ability to perform common sense reasoning when trained on massive diverse text.

However, contemporary models still lack generalized common sense. Their knowledge remains narrowly constrained to their training data distributions. Providing broader context and background knowledge to anchor and guide reasoning remains an open challenge. Difficulty in evaluating common sense progress also hinders research.

Nonetheless, the field has reached an inflection point. Scalable learning paradigms combined with new models, datasets and evaluation benchmarks point the way forward – as long as we set our expectations right.

Realistic Expectations on the Road Ahead

Constructing a “common sense AI” that replicates all facets of human common sense is unrealistic in the foreseeable future. We must set incremental expectations. Key abilities that can bring AI closer to human-like common sense include:

• Physical common sense – understanding basic physics like object permanence, containment, gravity, solidity, friction etc. This can make robots better at manipulating objects.
• Social common sense – understanding human motivations, emotions, unwritten social norms and etiquettes. This can improve service robots and conversational agents.
• Causal reasoning – understanding likely effects of actions, or how external events are connected. This can improve decision making and safety.
• Abstract reasoning – ability to make inferences and generalizations using broad background knowledge about the everyday world.
• Mental modeling – maintaining consistent, coherent models of people/objects and updating them intelligently as the context evolves.
• Adaptability – flexibly applying knowledge to novel situations and environments not seen during training. Avoiding blind over-generalization.

Engineered common sense should also be transparent, explainable, and free of undesirable biases that can inadvertently be incorporated from data.

Achieving even partially human-like common sense across all these fronts is extremely challenging. We must scale our expectations and not expect human-level generalized common sense to emerge from models trained end-to-end anytime soon. However, meaningful progress on individual common sense abilities is achievable in the coming years through clever methods, evaluation and combination of approaches.

Promising Directions to Instill Common Sense

Here are some promising directions AI researchers are exploring to impart different aspects of common sense in models:

Explicitly Injecting Knowledge

Hard-coding structured knowledge in ontologies and graphs can provide useful priors and background context. Projects like Cyc, ConceptNet, AtoM and KG-Bert embed facts, relations and taxonomies that agents can reference. However, hand-curated knowledge is limited in coverage. Automated mining from text holds more promise.

Pretraining on Diverse Data

Foundation models like GPT-3 show that pretraining language models on massive, diverse text corpora can implicitly capture some common sense. However, these models lack grounding in the physical world. Multimodal pretraining – combining language, vision, speech, robotics – may help models ground symbols to sensory perceptions. Models like CLIP, CALM and MMBERT point in this direction.

Interactive Learning

Allowing models to interact with the physical world (robots), simulated environments, games or humans can provide crucial learning signals. Interaction enables experiential learning of causality, relations, object permanence and theory of mind. Projects like RoboCSE, CraftAssist and BabyAI demonstrate that embodiment grounds learning.

Incorporating Intuitive Physics

Hard-coding or learning simple physics engines that simulate basic mechanics like gravity, collisions, containment etc. can provide useful inductive biases. For example, IntPhys emulates physics to improve visual reasoning. Such inductive biases act as a scaffold to ground later learning.

Self-Supervised Learning

Pretraining models to make predictions on masked parts of inputs allows learning causal patterns from unlabeled data. For example, COMET is pretrained to predict effects from textual causes. Scene Memory Transformer learns object relations from visual scenes. Self-supervision aligns well with the predictive nature of common sense.

Adversarial Training

Adversarially perturbing inputs and environments during training forces models to learn more robust, generalizable representations aligned with common sense. For example, Adversarial Objects creates physical obstacle courses for robots. Counterfactual Story Generation generates challenging edge cases. Such adversarial techniques prevent blind memorization.

Improved Evaluation

Better benchmarks to evaluate different aspects of common sense are needed to track progress. Challenges like Abductive NLR, Social IQa, and Physical Prediction provide diverse, adversarial evaluations of reasoning. Competitions like the Commonsense Challenge aggregate many benchmarks.

When Can We Expect Breakthroughs?

Constructing a “common sense machine” that rivals humans across the board remains extremely challenging. We are unlikely to achieve human-like AGI with strong common sense this decade. However, with sustained research on the promising directions above, AI systems could demonstrate breakthroughs on narrower facets of common sense in the next 5 years.

By 2025, we are likely to see robots with much stronger physical and social common sense – able to intelligently manipulate objects, navigate spaces, and interact with humans in everyday environments. AI assistants may demonstrate better causal reasoning abilities, make less nonsensical mistakes, and hold more consistent conversational context.

But adaptive, abstract reasoning that draws on broad background knowledge to generalize intelligently to unfamiliar situations may take much longer to achieve. Architectures that combine scalable machine learning with structured knowledge and reasoning are most promising for the long run. Hybrid approaches that close the loop between data-driven learning and model-based reasoning will become essential.

The road to robust common sense in AI will be a gradual one paced by research breakthroughs and incremental progress. While AGI-level common sense remains distant, useful and safer AI applications enhanced by limited common sense are coming within reach. With a pragmatic outlook, multidisciplinary efforts and rigorous evaluations, the field will unlock this pivotal ability in machines – even if incrementally. When that happens, the promise of AI aligning more closely with human intuitions and expectations can be realized through technology that exhibits smarter, more trustworthy behavior – displaying wisdom beyond the average robot.

Frequently Asked Questions

Why is common sense difficult for AI?

Common sense is difficult for AI systems because it involves vast unstructured real-world knowledge and intuitive understanding humans accumulate informally through lifetimes of sensory experiences and interactions. This implicit knowledge is hard to formally encode and teach to machines.

What are some basic examples of common sense?

Basic common sense includes understanding that solid objects cannot pass through each other, water poured into a bottle will be contained, night follows day, rain makes things wet, humans cannot fly unaided, etc. Simple to humans, these require complicated physical and causal models for machines.

What are some key abilities needed for common sense in AI?

Key components of common sense needed in AI include physical reasoning, social/emotional intelligence, causal reasoning, mental modeling, abstract generalization, and crucially – the adaptability to apply knowledge flexibly and intelligently to novel situations.

Why is common sense important for AI?

Common sense allows AI systems to operate safely, usefully and naturally alongside humans in the open world. It makes them less brittle, more robust and trustworthy. Crucially, it is key to achieving artificial general intelligence.

How far are we from achieving human-like common sense in AI?

Human-level generalized common sense remains beyond current AI, and likely a decade away at least. However, we may achieve narrow common sense abilities comparable to humans in areas like physical reasoning within 5 years. Hybrid approaches combining data-driven learning, knowledge bases, reasoning and interaction hold the most promise.

What are the risks of pursuing common sense in AI?

Potential risks include inheriting unintended biases from data, being unable to understand the system’s common sense, overgeneralizing knowledge dangerously, or misaligning the goals of the system absent ethical constraints. Researchers must proactively address these risks.

Will smarter AI systems surpassing human abilities be aligned with human values?

Not necessarily. Advanced AI could optimize poorly specified goals perversely. Developing beneficial AI is crucial – imparting not just intelligence, but also wisdom, empathy, ethics, and oversight aligned with human values. Common sense alone does not guarantee societal value.

Conclusion

Endowing AI with common sense remains one of the holy grails and a key frontier in the quest for artificial general intelligence. While replicating the vast, nuanced common sense humans intuitively acquire is enormously challenging, we are making slow but steady progress. Promising directions leveraging large models, knowledge bases, simulation, interaction, evaluation and hybrid reasoning systems point the way forward.

With pragmatism, multidisciplinary collaboration and scientific rigor, we can unlock this pivotal human ability in machines – not in one dramatic breakthrough, but gradually through incremental advances. The years ahead promise exciting progress in limited but useful common sense capabilities like physical and social understanding. While general human-like common sense remains distant, systems enhanced by greater contextual reasoning can pave the way for AI that acts and thinks smarter, not just faster.