At Rise of AI 2019 in Berlin, Helmut Linde, Global Head of Data Science at Merck KGaA, argues that the path toward Strong AI runs through solving the problem of invariant representations. He compares today's algorithms with biological intelligence, noting that a small child can recognize a cat from only a few examples, that a person recognizes a song regardless of which instrument or key it is played in, and that a story is recognized whether it appears as a film, an audio play or a book. Brains, he argues, can form invariant representations of real-world entities across very different sensory inputs, something current machine learning algorithms cannot yet do. He states that this gap, rather than raw performance on narrow benchmarks, is the key obstacle to progress toward Strong AI, and that the brain can serve as an inspiration for a next generation of learning algorithms. Although there is still very limited understanding of the brain's underlying computational principles, Linde argues there is reason to hope this biological algorithm can be reverse-engineered, which he says requires a genuinely interdisciplinary effort spanning neuroscience, computer science, mathematics and physics.
Key takeaways
- Linde frames the gap between biological and artificial intelligence as the ability to form invariant representations of real-world entities.
- A child can recognize a cat from only a few examples, which Linde contrasts with how current algorithms learn.
- People recognize a familiar song across different instruments or keys, illustrating invariance that current algorithms lack.
- A story is recognized as the same story whether experienced as film, audio play or book, another example of invariant representation.
- Linde argues that closing the invariant-representation gap, not just scaling current methods, is key to progress toward Strong AI.
Detailed notes
Why invariant representations are the key problem
Linde's central argument, as stated in the talk description, is that biological brains can form invariant representations of real-world entities, recognizing a cat after only a few examples, a familiar song played on a different instrument or in a different key, or a familiar story presented as a film, an audio play or a book, while today's machine learning algorithms lack this ability. He treats closing this gap as the key obstacle on the path toward Strong AI, rather than further scaling of current approaches.
An interdisciplinary path to reverse-engineering the brain
Linde argues that although there is currently very limited understanding of how the brain works conceptually, meaning what its underlying "algorithm" actually is, there is reason to hope this algorithm can eventually be reverse-engineered and implemented in a computer. He states that making progress on this question requires a highly interdisciplinary approach drawing on neuroscience, computer science, mathematics and physics together, rather than any one field working alone.
Questions answered in this talk
What does Helmut Linde identify as the key obstacle to Strong AI?
According to Helmut Linde at Rise of AI 2019, the key obstacle is that today's algorithms cannot form the kind of invariant representations of real-world entities that biological brains form, such as recognizing a cat from a few examples or a song across different instruments.
Why does Helmut Linde think the brain can inspire future AI algorithms?
Linde argues at Rise of AI 2019 that despite limited current understanding of how the brain works conceptually, there is reason to hope its underlying computational principles can be reverse-engineered and implemented in a computer, making the brain a useful inspiration for a next generation of learning algorithms.
What kind of research does Helmut Linde say is needed to progress toward Strong AI?
At Rise of AI 2019, Linde states that progress requires a highly interdisciplinary approach combining neuroscience, computer science, mathematics and physics, rather than advances in any single field alone.