← Back to list
Lex Fridman PodcastPodcast3 Oct 2019Source: lexfridman.comHost: Lex Fridman

Gary Marcus: Toward a Hybrid of Deep Learning and Symbolic AI

In plain words

AI expert Gary Marcus argues that deep learning (AI that learns from data) has big limits: it finds correlations but lacks common sense, like 'a bottle holds water' or 'people drink when thirsty.' He wants to combine deep learning with symbolic AI (which uses logic rules). Marcus is cautious about AI progress, saying it's gradual, not a sudden singularity. Key examples: DeepMind (lost $530M in 2018, its tech fails on language), GPT-2 (writes fluently but fails on simple logic), and Google Duplex (only handles simple tasks like booking haircuts, not general).

AI SummaryAI-generated · may contain errors · verify against the original

Gary Marcus, in his discussion on the Lex Fridman podcast, explored the necessity of integrating deep learning with symbolic AI. He pointed out that current deep learning suffers from fundamental limitations such as low data efficiency and a lack of common-sense reasoning, making it incapable of ach

~9 min full read · 9 sections
Deep Analysis

Gary Marcus: Toward a Hybrid of Deep Learning and Symbolic AI

At a Glance

Gary Marcus is a Professor Emeritus at New York University and founder of Robust.AI, who has long taken a critical perspective on the limitations of deep learning. The core theme of this issue: Current deep learning faces fundamental limitations in data efficiency, common sense reasoning, and causal understanding that cannot be resolved simply by scaling up computational power. A hybrid system that integrates symbolic AI with deep learning must be built. Marcus's core judgment: Deep learning excels at perception and classification, but it cannot grasp common sense such as "a bottle holds water" or "a person drinks water because they are thirsty"—and this is the "rate-limiting step" toward achieving general intelligence.


I. The Singularity Will Not "Arrive Suddenly": Intelligence Is a Multidimensional Variable

Marcus argues that the so-called "technological singularity" will not be a sudden shift at a single point in time, but rather a gradual, multidimensional evolution.

  • Intelligence is not a single IQ score: Marcus points out that intelligence encompasses multiple dimensions, including language, motor skills, and mathematics. Machines have already far surpassed humans in mathematics and games, but "lag behind my five-year-old child" in natural language understanding. He predicts that AI progress will be linear, rather than exponential as Ray Kurzweil believes.
  • Common sense is the "rate-limiting step": Machines lack common sense—they do not understand basic facts such as that a bottle holds water, that people drink water to quench thirst, or that dehydration is harmful. This limits reading comprehension (stories rely on background knowledge) and flexible problem-solving. Marcus emphasizes: "If we had a machine-interpretable common sense knowledge base, many things would accelerate relatively quickly."

Inference: Marcus believes that physical common sense (e.g., a key opens a lock) may be easier for machines to master than psychological common sense (e.g., understanding others' motives), because robots can acquire physical data through experimentation, whereas psychological experiments are constrained by ethical limitations.


II. The Fundamental Limitation of Deep Learning: Correlation ≠ Understanding

Marcus systematically critiques 10 challenges of current deep learning architectures, with the core issue being that systems learn correlations rather than causality and abstract concepts.

  • Closed Problems vs. Open Problems: Chess and Go are "closed problems"—with fixed rules and limited action spaces. In reading comprehension, however, "the next sentence could be anything," presenting a fundamentally different challenge. Marcus notes: "You cannot directly apply AlphaGo's technology to reading. DeepMind tried, and the results were poor."
  • Abstract Concepts Cannot "Emerge": Marcus cites a 1998 experiment as an example—neural networks could learn all even numbers but failed to generalize to odd numbers. The reason is that the system learns correlations between input and output nodes, rather than algebraic operations on variables. He quotes Bertrand Russell's metaphor: deep learning "has all the advantages of theft over honest toil."
  • Convolution as a Success Case of "Innate Programming": Marcus points out that convolution (the core of CNNs) is an abstraction manually programmed by Yann LeCun—enabling the system to recognize the same object at different positions. This precisely demonstrates that "we need more such innate features, not fewer."

Data Support: Marcus mentions that DeepMind posted a loss of $530 million in 2018, and despite massive investments in applying similar technologies to language, "productivity is far lower."


3. Hybrid Systems: The Inevitable Fusion of Symbolic AI and Deep Learning

Marcus explicitly opposes the extreme paths of "pure deep learning" or "pure symbolic AI," advocating for hybrid systems that combine the strengths of both.

  • The value of symbolic AI: Marcus emphasizes that variables, operations, and conditional logic (e.g., "if X > Y then execute Z") form the foundation of modern computing. He asks rhetorically, "Would you build an operating system or a browser using machine learning? No one would do that."
  • Lessons from expert systems: Marcus clarifies that he is not advocating for a return to "expert systems"—the expert systems of the 1980s were almost entirely hand-coded rules with no machine learning. What he advocates for is "new hybrid technology," leveraging deep learning's perceptual classification capabilities alongside symbolic systems' reasoning abilities.
  • AlphaGo is already a hybrid system: Marcus points out that AlphaGo is often described as a "deep learning system," but more accurately, it is a hybrid of deep learning and Monte Carlo tree search (a classic AI technique). He believes that in the future, people will "relabel hybrid systems as deep learning."

Implication: Marcus argues that symbolic AI may need its own "ImageNet moment"—a breakthrough demonstration that captures the world's imagination. The Allen Institute for AI (AI2) is building a commonsense reasoning benchmark dataset, which could serve as a catalyst.


4. Innate Knowledge: Drawing Design Inspiration from Biology

Marcus argues that the way human children learn—an innate knowledge framework combined with experiential learning—provides key design principles for AI.

  • The Necessity of Innate Knowledge: Marcus points out that children are born with "framework knowledge" about space, time, other agents, and causality. This is not a binary of "nature vs. nurture," but rather "nature and nurture must work together."
  • Evolution is Cumulative: Using the example of a baby antelope being able to walk down a mountain within hours of birth, Marcus illustrates that organisms carry a vast amount of innate knowledge. Evolution is inefficient (taking a billion years to develop the vertebrate brain plan), but once a "good idea" emerges, it spreads rapidly through the gene pool. He suggests that AI should draw from "biological bionics"—studying cognitive science, psychology, and linguistics.
  • Humans as a "Low Bar": Marcus agrees with Nobel laureate Danny Kahneman that humans are "a low bar that is easy to surpass." Machines should learn what humans excel at (flexible reasoning, language understanding) while avoiding human flaws (motivated reasoning, confirmation bias, poor memory).

Falsification Condition: If a future system can robustly predict "which containers will leak and which will not" simply by watching videos, Marcus says, "I would be delighted to see it."


5. The Prerequisite for Trustworthy AI: From Deep Learning to "Deep Understanding"

Marcus argues that the fundamental reason current AI is untrustworthy is that it only processes correlations without understanding abstract concepts such as "harm" or "intent."

  • Unalignable systems: Marcus points out that you cannot achieve "value alignment" with a system that only processes pixel correlations. Asimov's First Law of Robotics, "First, do no harm," requires the machine to understand what "harm" is—this demands abstract concepts, which deep learning handles poorly.
  • The translation gap: Even if an ethics committee formulates rules, someone must translate them into code or neural networks. Marcus emphasizes: "We don't know how to do that right now. If we can't even do it with Python or TensorFlow, we are deluding ourselves into thinking we can build trustworthy AI."
  • The six-question test: In his book, Marcus proposes six questions for evaluating AI reports, with the core being "demand a demonstration." If Sundar Pichai claims Google Duplex can converse like a human, you should ask, "Can I try it? How general is it?"

Extrapolation: Marcus argues that if AI could pass a "comprehension test"—for example, after watching Spartacus, answering "Why did everyone stand up and say 'I am Spartacus'?"—he would acknowledge that AI has made genuine progress.


Mentioned Positions

Position Analyst View Key Data
DeepMind Risk Warning 2018 loss of $530 million; poor results when applying AlphaGo technology to language tasks
GPT-2 Risk Warning Generates fluent text but lacks consistent concept understanding; fails on reasoning tasks like "DAX is DAX"
Google Duplex Risk Warning Only handles barber shop and restaurant reservations, and business hours inquiries; not general-purpose
CYC Neutral (Historical Case) 30 years of manually coded common sense, unsuccessful; Marcus argues this should not discredit the symbolic AI approach
AlphaGo/AlphaZero Neutral (Hybrid System Case) Deep learning + Monte Carlo tree search; not pure deep learning

Judgments Worth Remembering

1. Marcus: Intelligence is a multidimensional variable, not a single IQ score. Machines far surpass humans in mathematics and games, but lag behind a five-year-old child in language comprehension—progress across different dimensions is asynchronous, and the singularity will not be a single moment.

2. Marcus: Deep learning excels at perceptual classification but cannot grasp common sense like "a bottle holds water." This is the "rate-limiting step" toward general intelligence, because reading comprehension, physical reasoning, and social interaction all depend on such background knowledge.

3. Marcus: Neural networks learn correlations, not algebraic operations on variables. A 1998 experiment showed that a network could learn all even numbers but failed to generalize to odd numbers—21 years later, GPT-2 similarly failed on "DAX is DAX" type reasoning, with the underlying architecture unchanged.

4. Marcus: AlphaGo is a hybrid system, not pure deep learning. It incorporates Monte Carlo tree search (a classical AI technique), and in the future, people will "relabel hybrid systems as deep learning."

5. Marcus: Humans are "a low bar that is easy to surpass." In agreement with Kahneman: machines should learn the flexible reasoning at which humans excel, while avoiding human flaws such as motivated reasoning, confirmation bias, and poor memory.

6. Marcus: The prerequisite for trustworthy AI is "deep understanding," not "deep learning." You cannot align values with a system that only processes correlations—understanding "harm" requires abstract concepts, which deep learning handles poorly.

7. Marcus: Evolution is cumulative—once a "good idea" emerges, the gene pool spreads it rapidly. A newborn antelope can descend a mountain within hours of birth, indicating that organisms carry a vast amount of innate knowledge; AI should draw on cognitive science rather than starting from scratch.

8. Marcus: The six-question test for evaluating AI reports—the core is "ask to see a demo." If Sundar Pichai says a system can converse like a human, you should ask, "Can I try it? How general is it?"