Bengio, Yoshua

deep learning pioneer; professor at University of Montreal; scientific director of Mila

Yoshua Bengio: Deep Learning

Lex Fridman Podcast · Oct 20, 2018

• Yoshua Bengio
• deep learning
• long-term credit assignment
• disentangled representations
• catastrophic forgetting
• distribution generalization
• AI existential risk
• algorithmic bias
• machine teaching
• model-based reinforcement learning

Core ideas

Bengio identifies long-timescale credit assignment as one of the most mysterious and potentially fruitful mismatches between biological and artificial neural networks: humans can revise their interpretation of something that happened a year ago based on new evidence, effectively performing credit assignment across essentially arbitrary time spans via episodic memory, while current recurrent architectures handle sequences of dozens to hundreds of timesteps reasonably well but degrade sharply beyond that. He connects this directly to the problem of selective forgetting — humans don't remember everything, they remember what matters — and speculates this connects meaningfully to open questions about consciousness and emotion (what determines which experiences become consciously accessible and durably stored).

On deep learning's current representational weaknesses, Bengio argues today's large-scale models trained on images or text achieve only a shallow, low-level understanding — not robust or abstract in the way human understanding is — and that the fix isn't simply architectural or dataset-driven but requires rethinking training objectives themselves: moving from passive observation of data toward active agents that learn by intervening in the world and discovering cause-and-effect relationships, and jointly learning from language and world-grounded experience together rather than separately, since language provides crucial clues about which high-level semantic concepts actually matter (citing evidence that purely unsupervised representation learning fails to produce representations as powerful as those guided even minimally by labels). He explicitly rejects that merely scaling network depth (e.g., from 100 to 10,000 layers) will solve these representational problems, arguing genuinely different training frameworks are needed — while separately agreeing that current models are likely also underparameterized relative to the brain's scale, and noting (as a silver lining for academic researchers without Google-scale compute) that state-of-the-art deep learning still fails to robustly understand even simple synthetic environments like grid worlds without needing far more examples than a human child would require for the same task.

Bengio frames classical symbolic/expert-system AI as having correctly identified important goals (representing common-sense knowledge, structured reasoning) that failed in practice partly because much human "common sense" knowledge isn't consciously accessible or codifiable into explicit rules, and partly because expert systems couldn't replicate the representational power of neural networks' distributed representations. He argues neural networks in turn lack a different kind of compositionality found in symbolic systems and language: while his long-standing "disentangled representations" research agenda (separating a representation's underlying causal factors into cleanly separable variables) remains valuable, he now believes it's insufficient on its own, since it's equally important to disentangle the relationships (rule-like mechanisms) between those variables from one another — analogous to how individual expert-system rules can be modified independently without breaking unrelated rules, in contrast to neural networks' vulnerability to catastrophic forgetting, where learning new things can destroy previously learned knowledge because everything is entangled together in one large parameter blob. He hypothesizes this kind of disentanglement (of both variables and their governing relationships) is only achievable in a proper high-level semantic representation space, not directly in raw sensory/pixel space, and that this factorized structure is precisely what would enable genuine generalization beyond the standard machine-learning assumption that test and training data share the same distribution.

He illustrates generalization beyond the training distribution with the example of reading a science fiction novel set on an alien planet with a visually alien surface — comprehensible because the same underlying physical laws and (often) causal/social relationships still transfer, even though the surface-level "distribution" looks entirely different — arguing humans routinely perform this kind of transfer by carrying forward knowledge of deep causal and physical mechanisms rather than surface statistics, a capability current machine learning has no principled account of.

On AI safety, Bengio draws a sharp distinction between two separate conversations: the AI research community's internal discussions and what he considers the far more urgent and useful public conversation, which should focus on near- and medium-term societal risks — surveillance ("Big Brother" facial recognition scenarios), autonomous weapons ("killer robots," which he's careful to distinguish from Terminator-style scenarios), labor market disruption, concentration of power, and discrimination/bias, several of which he considers genuine threats to democratic institutions. He considers existential risk from runaway superintelligence a legitimate but currently low-priority topic worth some academic study (comparing it to studying asteroid-impact scenarios), stating it's implausible given his understanding of how current neural networks and machine learning actually work, while acknowledging he can't rule out very different future AI paradigms decades out. Asked about the film Ex Machina, Bengio criticizes its portrayal of science as a fundamentally false picture — real scientific progress happens incrementally through broad collaboration and community interaction, information inevitably flows and leaks even within corporate labs, rather than through a single isolated genius working in secret — and considers this a legitimately harmful public misconception, since it distorts how people understand real science, while affirming that diversity of research direction (people pursuing orthogonal or even contradictory approaches to his own) is essential and healthy for the field rather than something to be discouraged.

On algorithmic bias, Bengio distinguishes near-term technical interventions (using adversarial training methods to make models less sensitive to protected variables, applied to known-biased datasets — techniques he considers mature enough that governments should already be regulating high-stakes industries like insurance to require them, since companies won't adopt them voluntarily given the accuracy costs involved) from a longer-term, more speculative research agenda: teaching machines to model human emotional and moral responses, starting with something as basic and cross-species as anger triggered by perceived injustice, potentially first in simplified virtual/game environments where agents' interactions can be labeled with likely human emotional reactions — a goal he frames as realistically achievable only in proportion to how well systems understand the broader world, which remains a long way off.

Bengio argues the machine learning community has under-invested in the deeper problem of "machine teaching" — not just data annotation as a rote task, but genuinely effective pedagogical strategies for how a teaching agent (potentially eventually human) can most efficiently guide a learning agent, citing his own group's "BabyAI" project (a teacher-learner game environment) as an early exploration of this direction, and noting its clear relevance as human-machine interactive learning becomes more prevalent. On natural language understanding, he identifies non-linguistic, world-grounded knowledge — illustrated by Winograd schema sentences requiring real-world causal understanding to disambiguate — as the genuinely hard remaining challenge, arguing this points toward systems that jointly model how the world works causally and how to express that understanding linguistically; he considers the underlying learning principles language-independent, expressing interest in systems that could learn equally well from any human language rather than English specifically.

Reflecting on surviving an AI winter, Bengio credits close friendships within the research community and a discipline of trusting one's own well-evidenced scientific intuition over prevailing fashion — while stressing this must remain responsive to actual experimental contradiction, not blind stubbornness — as what allowed him to persist through years when his research directions were unfashionable before eventually bearing fruit. He argues celebrated "seminal moments" in AI history (like AlphaGo) are somewhat overrated relative to the actual gradual, incremental nature of scientific progress — a single small technical step can have outsized commercial or public impact once it crosses a threshold enabling a previously infeasible application, even though the underlying science advances gradually — and identifies reinforcement learning/agent learning and generative models (GANs and related approaches) as the current areas of greatest excitement in his own institute, arguing that moving from purely policy-gradient-based reinforcement learning (which never explicitly models the world) toward genuine model-based RL — capturing underlying causal mechanisms rather than just direct policy mappings — is essential for building agents that generalize well to new distributions, even though this commercial payoff (unlike AlphaGo's fame) hasn't yet materialized.

Excerpts