Interview, Fireside Chat, Conference Presentation
Gary Marcus: Limits of Deep Learning | AI Podcast Clips
- Core Thesis: Deep learning systems currently rely on statistical approximations of correlations rather than "cognitive models" of reality, lacking true understanding of everyday phenomena and common sense.
- Systems fail when environmental variables change slightly (e.g., moving a paddle in a video game) because they lack robust representations of objects like "balls," "walls," or "paddles."
- Solving AI's fundamental challenges requires moving beyond pixel-level correlation to building systems that understand causal mechanisms (e.g., why bottle threads seal with caps).
- Limitations of Pure Data-Driven Approaches:
- The author argues that relying solely on training deep neural networks on massive video datasets will not yield robust common sense.
- No system has yet been built that can robustly predict physical outcomes (e.g., container leakage) purely by watching videos without explicit modeling.
- Deep learning systems learn complex correlations between input and output nodes but do not perform structured operations over variables (algebraic reasoning).
- Historical experiments (1998) showed auto-associator networks failing to generalize to odd numbers after training only on even numbers, as they tracked output probabilities rather than learning the underlying mathematical rule.
- Unlike humans who apply algebraic operations (e.g., $f(x) = y + 2$) to new inputs, neural networks emulate these patterns without grasping the variable-based logic.
- The Role of Symbolic AI and Engineering:
- Common sense requires tools from "good old-fashioned AI," including symbols, logic, and explicit programming, which are not obsolete but necessary complements to deep learning.
- Human experts must continue to perform "hard engineering work" to encode complex knowledge and design specifications into systems.
- Example: A 45-page paper by the author and Ernie Davis required three months of manual logical formulation just to define the concept of a "container."
- Abstractions like convolution are currently "innate" features programmed by humans (as seen in Jan LeCun's work) rather than learned by the system.
- The author advocates for adding more innate, human-engineered abstractions, not fewer.
- Symbolic manipulation (conditionals, variable comparisons) is essential for domains requiring absolute reliability, such as operating systems or web browsers, which would be unsafe if built solely on probabilistic machine learning.
- Future Directions and Methodology:
- Active Learning: A preferred hybrid approach involves continuous interaction where deep learning extracts patterns from data while humans guide the learning framework and procedures.
- Differentiable Programming: Emerging tentative work in this area may allow systems to incorporate structured operations over variables, though it is not yet standard in models like those used for ImageNet.
- Critique of "End-to-End" Learning: The author rejects the notion that machine learning can replace all software engineering (e.g., the fantasy that computer programmers are obsolete), citing the economic reality that developers still rely heavily on symbolic code.
- Causal Reasoning: Moving from correlation to causality is identified as a prerequisite for building robust systems that function in the real world.
- Disagreement with Current Trends:
- The author disputes claims by figures like Yann LeCun and Geoffrey Hinton that unsupervised learning or deep learning alone can replace the need for symbolic representation.
- The view that "deep learning is the only path forward" is characterized as destructive and incorrect, given that current architectures cannot reify abstract concepts or handle the complexity of physical world models.
- The author notes that while Hinton suggests throwing out symbols, he remains evasive on what exactly needs to be discarded, despite the author's stance that variables and operations over them are fundamental to computing and AI.