In a world dominated by neural networks like ChatGPT, Amazon is adopting a hybrid approach called neuro-symbolic AI, which powers the Rufus shopping assistant and advanced warehouse robots, solving problems that neural networks alone cannot.
Article contents:
- Introduction: Beyond neural networks
- The problem with large language models: Hallucination and lack of trust
- What is neuro-symbolic AI?
- Amazon in action: Vulcan robots and the Rufus assistant
- How it works: From mathematical logic to code
- Combating hallucinations with 99% accuracy
- The great debate in the AI world
- Conclusion: The moment for neuro-symbolic AI has arrived
Introduction: Beyond neural networks
Neural networks are not the only game in town when it comes to artificial intelligence, though you could be forgiven for thinking otherwise after the boom sparked by the arrival of ChatGPT in 2022. But now, a hybrid approach called neuro-symbolic AI is gaining momentum with a handful of companies, most notably Amazon, as developers push AI far beyond the comfort zone of neural networks.
The problem with large language models: Hallucination and lack of trust
ChatGPT and other generative AI tools can predict the next word in a sentence remarkably well, but they cannot always determine whether something is true when there are only limited examples to rely on. They can deliver unusually incorrect answers with irrational confidence. That is unacceptable when models begin taking actions on behalf of people or businesses.
“It’s not a problem that can be solved simply by throwing more compute power at it,” says Grant Passmore, co-founder and co-CEO of Imandra, an AI startup. “Neuro-symbolic AI is the solution because it combines the statistical modeling of large language models with the intuitive, logical approach of symbolic reasoning.”
What is neuro-symbolic AI?
Instead of the data-driven statistical operations of neural networks, symbolic reasoning uses logic and symbols to solve problems that can be expressed in code. Neuro-symbolic AI is a hybrid approach that combines the power of both worlds: the statistical capability of neural networks and the logical precision of symbolic reasoning. This approach has another advantage: cost. Unlike pure LLM deployments that require clusters of expensive GPUs, neuro-symbolic agents split their workloads between GPUs and standard CPUs.
Amazon in action: Vulcan robots and the Rufus assistant
Amazon has about 20 teams building automated reasoning combined with other technologies for various uses across the company. For example, the new Vulcan warehouse robots use automated reasoning for spatial problems like determining the optimal place to pick up or place an item. Meanwhile, neural networks enable them to handle perception tasks including image classification.
Automated reasoning also provides the foundation for Rufus, the neuro-symbolic AI shopping assistant that Amazon made available last year. The bot uses a large language model to help customers navigate a wide range of products, and relies on automated reasoning to make its responses more relevant and avoid errors.
How it works: From mathematical logic to code
“We know the Pythagorean theorem is true, even though we cannot count all possible triangles,” says Byron Cook, vice president and distinguished scientist at Amazon. “But mathematical logic allows for a bounded proof, where each step is verifiable.” This same principle is applied to code, where code is treated as a formula to prove its correctness against desired properties.
Combating hallucinations with 99% accuracy
Amazon announced a new neuro-symbolic feature called “automated reasoning checks” that reduces AI hallucinations and identifies correct model responses with up to 99% accuracy. Amazon said it is the first and only generative AI guardrail to do so. It can be used to formalize truth in specific domains like healthcare or return policies.
The great debate in the AI world
Competing approaches over the years have led to sharp public debate about the nature of intelligence. AI researcher and cognitive scientist Gary Marcus has long argued that the limits of LLMs must be offset by neuro-symbolic models. While Marcus says today’s hybrid systems still need to integrate innovations, it is a major step forward after decades of viewing AI as two separate branches.
Conclusion: The moment for neuro-symbolic AI has arrived
Marcus wrote in an article last month: “Neither neural networks nor classical AI can truly stand alone. We must find ways to combine them. After a 30-year journey, I believe the moment for neuro-symbolic AI has finally arrived.” With companies like Amazon embracing this approach to solve real-world problems, that view appears to be gaining real momentum.
