Ian Goodfellow on the Limits of Deep Learning, GANs, and the Future of AI
This article summarizes a conversation with Ian Goodfellow, a leading figure in deep learning. He's known for his deep learning textbook and for coining the term Generative Adversarial Networks (GANs). The discussion spans the limitations of deep learning, the nature of neural networks, GANs, and the future of AI. This conversation was part of the Artificial Intelligence podcast with Lex Fridman.
The Limits of Deep Learning
Goodfellow opens by discussing the limitations of deep learning, particularly its reliance on large amounts of data, especially labeled data. While unsupervised and semisupervised learning algorithms help, they still require substantial unlabeled data. Reinforcement learning, though not needing labels, requires extensive experiences.
The need for better generalization capabilities is a significant bottleneck. Goodfellow also emphasizes that deep learning is typically a submodule within a larger system, like AlphaGo's value function estimation or reinforcement learning algorithms' action selection modules.
Neural Networks as Programs
Goodfellow considers neural networks as a type of program. He views deep learning as learning programs with multiple steps. The depth of a TensorFlow graph represents the number of sequential steps, while the width signifies parallel steps. He argues that the key innovation of deep learning is the ability to have sequential program steps.
He contrasts this with earlier machine learning approaches like support vector machines, where operations were largely parallel. He sees deep learning more as iterative refinement of a representation (a state of understanding) than a singular representation building process.
Cognition, Consciousness, and Emergence
The conversation briefly touches on philosophical questions about cognition and consciousness. Goodfellow believes cognition can emerge from sequential representation learning. He finds consciousness harder to define, particularly regarding qualitative states of experience (qualia). He acknowledges the difficulty in formalizing or experimentally determining consciousness in AI systems. He is more optimistic about selfawareness emerging from current architectures, particularly in reinforcement learning, which forces agents to model their effect on the environment.
The Role of Computation and Data
Goodfellow is optimistic that increased computation and the right kinds of data can lead to significant advancements in AI, bridging the gap between limited and humanlevel cognition. He emphasizes the importance of multimodal data, integrating various senses and experiences, similar to how the human brain operates.
Adversarial Examples: Security Liability or Tool for Improvement?
Goodfellow's thinking on adversarial examples has evolved. Initially, he saw them as revealing fundamental problems with machine learning. Now, he views them primarily as a security liability. There appears to be a tradeoff between accuracy on adversarial examples and accuracy on clean examples.
While training against weak adversarial examples can improve accuracy on clean data (as seen with MNIST), this doesn't consistently hold up on other datasets or against stronger adversaries.
GANs: Generative Adversarial Networks
Goodfellow recounts the story of conceiving GANs during a bar conversation, overcoming initial skepticism from friends. He attributes the breakthrough partially to lowered inhibitions from alcohol. The initial skepticism stemmed from the difficulty of training two neural nets simultaneously.
GANs are a type of generative model that uses a twoplayer game where a generator tries to create realistic data (e.g., images), and a discriminator tries to distinguish between real and fake data. This adversarial process leads to the generator becoming better at producing realistic data. While the theory suggests GANs converge to a Nash equilibrium, the practical reasons for their success in generating realistic images remain somewhat mysterious.
GANs Beyond Image Generation
The conversation explores applications of GANs beyond image generation, including semisupervised learning (using the discriminator as a classifier) and domain adaptation (using adversarial techniques to extract features that are invariant across different domains).
GANs and Fairness
Goodfellow discusses how adversarial machine learning can help models be more fair by preventing them from using sensitive variables (e.g., gender) in their predictions. Another potential use is cycleGANs to transform data from one demographic group to another, allowing researchers to audit potential system discrimination.
Deepfakes and Authentication
Goodfellow expresses less concern about deepfakes in the long term (20 years) but acknowledges a potentially bumpy cultural transition in the near future. He believes authentication mechanisms will eventually emerge to verify the authenticity of content. He envisions systems where devices cryptographically sign content, making it possible to verify its origin.
Future Challenges and the Path to AGI
Goodfellow identifies fairness and interpretability as ripe areas for new ideas in deep learning. He believes that better definitions of these concepts are crucial, even without immediate algorithmic breakthroughs. He also sees simulation and a lot of computation as key ingredients for building systems with humanlevel intelligence (AGI).
He believes AGI requires environments that have a wide diversity of experiences and he is optimistic about the idea of an agent that is able to download, extract, and understand the data of different AI tasks with little to no assistance by a human engineer. He sees resistance to adversarial examples as one of the most important security challenges for researchers to solve.