Yann LeCun on Benchmarks, Interactive Environments, and Why Human Intelligence Is Not General

The host recalls LeCun's advice not to be fooled by people who claim to have a solution to artificial general intelligence, or to have figured out how the brain works. LeCun's test is to ask for their error rate on MNIST or ImageNet, and the host jokes that this is a little dated. The idea is that benchmarks and practical testing are where ideas actually get tested.

LeCun clarifies that a benchmark may not be completely practical, and it can be a toy dataset, as long as it is a task the community has accepted as a standard. He gives the bAbI tasks, proposed at FAIR years earlier, as a toy problem for testing reasoning and access to working memory that proved useful. What struck him was how many people with money to invest were taken in by pitches such as having the algorithm of the cortex in return for $50 million.

The host notes that new ideas may not have benchmarks yet, and asks what benchmarks might move toward reasoning. LeCun says many people are working on interactive environments for training and testing intelligent systems. The classical supervised paradigm splits data into training, validation and test sets, assuming samples are statistically independent and their order does not matter.

That fails when the machine's answer or action decides the next sample, as in robotics. A robot moves to a new room, and where it goes changes what it sees, so samples depend on each other, and moving through space means the next sample probably comes from the same building. All the assumptions behind the train-test hypothesis break once a machine can act on the world.

LeCun says this is where the field is heading. People are setting up artificial environments, such as a robot running around a 3D model of a house and interacting with objects, simulated robotics like OpenAI Gym or MuJoCo, and games.

LeCun says he does not like the term AGI because it implies human intelligence is general, when it is very specialized and we only like to think of ourselves as general. The host counters that what is impressive is the ability to learn across many domains and integrate that knowledge. LeCun answers with a quasi-mathematical demonstration about the roughly one million fibers in one optic nerve.

If those fibers were binary, the visual cortex would receive one million bits. If you cut the nerve and inserted a device that applied a fixed random permutation, LeCun says there is no way the visual cortex would learn vision to the same quality, even if done in infancy. Neurons there only connect locally, so pixels that are neighbors in the world would land far apart, which shows the hardware is built to support the locality of the real world.

He then counts: with one million binary inputs there are 2 to the 1 million input combinations, and so 2 to the 2 to the 1 million Boolean functions. Only a tiny sliver can be computed by the visual cortex. The host objects that the brain adapts impressively, but LeCun says we feel general because we only consider things we can apprehend.

He compares this to physics. For a gas we know pressure and temperature, which is a tiny number of bits compared with every molecule's position and momentum. What we cannot know we call heat or entropy, and there may be strong structure in it that we are not wired to perceive. He ties it to Kolmogorov complexity, where every bit string is random except those you can write down, and suggests simply calling it artificial intelligence, with human-level meaning "damn impressive".