Artificial intelligence has advanced tremendously in recent years, with machines capable of beating humans at complex games like chess and Go. However, as impressive as these feats are, AI still lacks fundamental human skills that we take for granted – things like common sense, social intelligence, and general knowledge about how the world works. This absence of basic reasoning and understanding is commonly referred to as the “common sense problem” in AI.
In this article, we’ll explore why common sense is so critical yet so elusive for AI, look at key areas where the lack of common sense manifests, review different approaches researchers are taking to impart common sense, and discuss what progress still needs to be made before AI can exhibit true human-level intelligence.
Why Common Sense is Critical for AI
Common sense encompasses all the basic facts, understandings, and reasoning skills that humans accumulate through life experience. This includes:
- Basic facts about the world – that water is wet, fire is hot, ice is cold.
- An understanding of how the physical world works – that objects fall down, liquid takes the shape of a container.
- Social norms and conventions – greetings, manners, appropriate conversational topics.
- The ability to make simple inferences and deductions based on existing knowledge. For example, if you walk into a dark room and turn on the light switch, you expect the room to become bright.
Humans are not born with common sense; we learn it gradually through daily interaction with the world and with other people. But this common sense knowledge is so ingrained that we take it completely for granted.
For AI systems, however, these basic assumptions and inferences do not come naturally. AI relies on rules, logic, and patterns derived from data. It does not have the same innate understanding of the world that allows humans to handle novel, unpredictable situations.
Without common sense, AI runs into problems:
- Brittle performance – AI systems excel at narrowly defined tasks under constrained conditions. But change the conditions even slightly, and the AI fails unpredictably because it does not understand the general properties and relationships that underlie the data.
- Inability to transfer knowledge – Humans can apply knowledge learned in one setting to totally different settings. Without an intuitive understanding of the world, AI systems cannot transfer knowledge between tasks.
- Lack of robustness – A small perturbation in the input data can lead an AI system to make completely wrong or nonsensical predictions. This demonstrates its fundamental lack of understanding.
- Inability to handle open-ended tasks – Real world situations are complex, nuanced, and unpredictable. With no common sense, AI systems cannot operate effectively except in simplified, limited environments.
For artificial intelligence to be integrated into the messy real world – controlling self-driving cars, having natural conversations, or performing open-ended office work – common sense is essential. That’s what makes imparting common sense into AI such an important challenge.
Where the Lack of Common Sense Manifests
The lack of basic common sense manifests in AI systems today across multiple different domains and tasks:
Language Models and Chatbots
AI systems that process or generate natural language, like large language models such as GPT-3, often demonstrate a lack of common sense when generating text. Some examples:
- Using absurd or contradictory statements – “The dog climbed up the tree and swam across the river”.
- Failing to differentiate fiction vs non-fiction – chatbots might talk about fictional characters as if they are real.
- Hallucinating implausible scenarios – “The astronaut removed his helmet on the moon and was able to breathe normally.”
- Getting basic facts wrong – “Paris is the capital of India.”
This happens because while these models are very good at continuing credible passages of text using patterns from data, they have no inherent concept of what is actually true or plausible.
Computer Vision Systems
Computer vision models that caption images or answer questions about images also stumble due to missing common sense:
- Mis-identifying objects, their attributes or relationships – labeling a dog as a cat.
- Failing to understand image context – identifying a person is “skiing” when image just shows skis leaning against a wall.
- Missing obvious details – not noticing the lack of wings on an “airplane”.
- Failing to identify illogical or contradictory elements – labeling a picture with both “daytime” and “night sky”.
Again, these systems lack the basic background knowledge about the world that allows humans to interpret images.
Robots that operate autonomously in uncontrolled real-world environments lean heavily on common sense:
- Navigating crowded spaces requires understanding social conventions and norms.
- Handling new objects or tools requires intuiting probable uses and properties.
- Processing language commands requires basic reasoning about implications.
Lacking this intuitive understanding of the world makes it challenging for robots to function beyond limited pre-programmed scenarios.
As we want to deploy AI systems in more expansive real-world roles, imbuing them with common sense becomes increasingly critical. While today’s AI excels at narrow tasks, common sense is key to general intelligence.
Approaches to Teaching Common Sense
Given the vast scope and subtle nuances of common sense knowledge, teaching it to AI systems poses a significant challenge. Here are some of the approaches being explored:
Curating Common Sense Databases
Many research efforts have focused on manually curating giant databases of common sense facts and rules, such as:
- Cyc – One of the longest-running projects, Cyc comprises over 250,000 common sense assertions like “Trees require water to survive”.
- ConceptNet – Contains over 21 million assertions mined from crowdsourced resources.
- ATOMIC – Focused on inferring likely pre- and post-conditions around human activities.
The hand-coded knowledge in these databases aims to supplement the statistical learning capabilities of AI. However, given the open-ended scope of common sense, these resources are still minuscule compared to what an average human knows.
Top 6 Forex EA & Indicator
Based on regulation, award recognition, mainstream credibility, and overwhelmingly positive client feedback, these six products stand out for their sterling reputations:
|Gold Miner Pro FX Scalper EA
|FXCore100 EA [UPDATED]
|Golden Deer Holy Grail Indicator
|Forex Trend Trading Course
|Forex Copy Trade
|Forex Fund Management
Learning from Examples
Rather than manually encoding common sense, we can also train AI models to learn from examples that demonstrate common sense reasoning.
- Natural language inference (NLI) – AI systems are trained on sentence pairs and must determine if one sentence semantically entails, contradicts or is neutral with respect to the other. For example, identifying that “The boy jumped into the lake” contradicts “The boy stayed dry”.
- Commonsense reasoning tasks – Models are trained to answer questions that require different types of commonsense inferencing about situations.
- Virtual environments – AI agents are placed in simulated interactive environments like video games to learn intuitive physics and relationships through experimentation.
By exposing models to diverse examples that encapsulate common sense, models can learn to make similar inferences about novel situations. But providing sufficient training data remains challenging.
Leveraging Unstructured Data
Rather than relying on specialized datasets, we can leverage the millions of natural language artifacts created by humans, like books, web pages, and social media, which implicitly reflect common sense.
Techniques like self-supervision allow AI models to harvest common sense patterns from these diverse corpora without explicit labeling. Large language models like GPT-3 are trained on massive text corpora in this self-supervised paradigm. However, these models tend to learn shallow statistical tendencies rather than robust generative knowledge. Combining self-supervision with other techniques may yield better results.
There is also interest in learning from non-verbal cues like images, videos, and human actions to instill better intuitive understanding. Multimodal training could strengthen common sense acquisition.
Overall, learning common sense from data still remains a wide open research question. Combining different data sources and learning paradigms will likely be needed.
The Road Ahead for Common Sense
While today’s AI systems still lack basic common sense, research interest and progress in this problem has grown tremendously in recent years. Here are some key directions needed to achieve human-like common sense in AI:
- Hybrid approaches – Combining rule-based common sense databases, supervised learning from examples, and self-supervised learning from diverse corpora will likely work better than any single approach.
- Richer representations – Rather than surface statistical tendencies, AI models need more structured, causal, hierarchical representations of knowledge.
- Interactive learning – Allowing AI agents to learn actively through interacting with environments, asking questions of humans, and getting human feedback.
- Shared foundations – Building common sense resources, architectures, and capabilities that transfer across different AI application domains.
- Long-term progress – Developing evaluation benchmarks to track progress over the long-term. Unlike narrow AI applications, common sense will require prolonged research spanning decades.
While AI has achieved superhuman performance in constrained environments, the quest to develop human-like common sense and reasoning abilities remains in its infancy. But given the foundational importance of common sense for achieving true general intelligence, expect this to be a very active area of AI research for decades to come. The day when we can have natural, wide-ranging conversations with AI akin to another human still seems far away. But equipping our AI assistants with more common sense to become less mystified by the human world is an important milestone towards that goal.
Frequently Asked Questions on Common Sense in AI
Here are answers to some frequently asked questions about the common sense problem in artificial intelligence:
Why is common sense difficult for AI systems to acquire?
There are a few key reasons why common sense remains difficult for AI systems:
- Scope – Common sense encompasses a vast body of knowledge accumulated through lifetimes of diverse experiences. It’s challenging to comprehensively encapsulate all this background knowledge.
- Nuance – Common sense often relates to subtle aspects of the world that are hard to codify into simple rules. Intuitions and unspoken assumptions are hard toCapture.
- Context-dependence – What constitutes common sense depends heavily on the situation and context. Flexibly applying common sense is hard for rigid AI systems.
- Grounding – Humans accumulate common sense through rich interaction with the real physical world from infancy. Most AI systems lack this kind of grounded experience.
Overall, both the breadth and subtlety of common sense knowledge poses difficulties for formal knowledge representation and statistical learning approaches used in AI.
What are some examples that illustrate the lack of common sense in today’s AI?
Here are a few examples that highlight the lack of common sense in current AI systems:
- Language models generating absurd statements like “The flowers sang a song while the sun was asleep”.
- Image captioning systems mis-identifying objects, attributes, or relationships in images.
- Chatbots making contradictory statements within a conversation.
- Robots having trouble understanding social norms and conventions to navigate crowded spaces.
- Self-driving cars getting confused by novel objects or situations not present in training data.
- AI systems making illogical inferences like “If John skipped breakfast, he must be hungry for supper”.
- Smart assistants providing nonsensical responses to queries requiring basic reasoning.
These examples demonstrate how AI systems can fail on simple situations that would be obvious to humans using common sense.
What are some key common sense capabilities missing in AI today?
Some major areas where today’s AI systems lack human-like common sense include:
- Intuitive understanding of objects, their interactions and properties.
- Basic physical and social reasoning skills.
- The ability to transfer knowledge and experiences between different contexts.
- Distinguishing factual information vs opinions vs speculation vs fiction.
- Handling ambiguous, implied or contradictory information appropriately.
- Making appropriate inductive generalizations and analogies based on experience.
- Understanding time, goals, intentions, causes and effects.
- Incorporating unspoken assumptions and norms shared by people.
Advancing AI to handle these more flexibly and robustly requires modeling the world in a way more grounded in everyday human experiences and mental models.
What are some promising ways to teach common sense to AI systems?
Some promising approaches to instill common sense in AI include:
- Using interactive environments like games and simulations to allow AI agents to learn common sense intuitively through experimentation.
- Training neural networks on datasets designed specifically to teach different aspects of common sense reasoning.
- Leveraging massive sets of narrated events, videos, dialogue, stories, and books to learn from examples of how humans apply common sense.
- Combining hand-crafted common sense knowledge bases and inference rules to complement data-driven machine learning.
- Allowing humans to provide interactive feedback and corrections to impart common sense knowledge.
- Training multimodal models that can learn associations between visual, textual and auditory inputs.
- Using transfer learning and modular architectures to share common sense capabilities between different AI systems.
A combination of techniques drawing upon curated knowledge bases, interactive learning, and absorptbotoing from large unlabeled corpora will likely be needed to effectively learn and represent common sense.
How far are we from developing human-like common sense in AI?
While progress is being made, AI systems today still have extremely limited common sense compared to humans. Some researchers estimate replicating the common sense reasoning abilities of a four-year-old child to be around a decade away.
However, AI may manifest common sense differently from humans based on its training. Matching the breadth, flexibility, contextual adaptability and grounded intuitiveness of human common sense is a very long-term challenge – likely requiring decades more of research and data to achieve.
But having AI systems exhibit even some rudimentary common sense that makes them less brittle and more robust when operating in the open world would still be a major milestone. Though human-like common sense is still a distant vision, equipping AI with stronger reasoning abilities step-by-step remains an important endeavor.
The absence of common sense remains a major barrier preventing artificial intelligence from achieving more flexible, generalizable, and trustworthy behavior when operating in the open real world. But given the centrality of common sense to human cognition, instilling these capabilities into our AI assistants emerges as an extremely important long-term research area, even if human-like common sense remains a distant target.
While rule-based and data-driven techniques have limits, combining curated knowledge with interactive grounded learning across modalities offers promise. With sustained research progress on benchmarking tasks, representation learning, and knowledge transfer, we can hope to see AI systems exhibit ever stronger semblance of common sense – even if anchored in the idiosyncrasies of machine intelligence rather than human mental models. But elucidating the complex web of concepts, assumptions, and inferences that comprise common sense marks an epic challenge on the road towards artificial general intelligence.
Top 10 Reputable Forex Brokers
Based on regulation, award recognition, mainstream credibility, and overwhelmingly positive client feedback, these ten brokers stand out for their sterling reputations:
|Open New Account
|MT4, MT5, RTrader
|Standard, Cent, Zero Spread
|Welcome Bonus $30
|Open RoboForex Account
|Standard, Cent, Zero Spread
|Top Forex Broker
|Open AvaTrade Account
|Standard, Cent, Zero Spread
|Open Exness Account
|ASIC, CySEC, FCA
|Standard, Micro, Zero Spread
|20% Deposit Bonus
|Open XM Account
|MT4, MT5, CTrader
|Standard, Zero Spread
|Best Paypal Broker
|Open ICMarkets Account
|ASIC, CySEC, FCA
|Standard, Zero Spread
|Best USA Broker
|Open XBTFX Account
|Standard, Micro, Zero Spread
|Welcome Bonus $50
|Open FXTM Account
|ASIC, CySEC, FCA
|Standard, Cent, Zero Spread
|100% Deposit Bonus
|Open FBS Account
|Best Crypto Broker
|Open Binance Account
|Best Trading Platform
|Open TradingView Account