将神经符号人工智能集成到LLM中以增强推理能力.pdf

1、 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Dr.Kanna ShimizuSenior Manager,Applied ScienceNeuro-Symbolic AI Lab,AWSNeurosymbolic AI Integratio

2、n in LLMs to enhance reasoningI N V 5 1 7 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Agenda1.Why do we need other computing techniques?2.What are other computing approaches?3.How do we combine the best of both worlds?4.What did we learn?2025,Amazon Web Services,Inc.or its aff

3、iliates.All rights reserved.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.1.Why do we need other computing techniques?2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Models are getting bigger and better 2025,Amazon Web Services,Inc.or its affiliates.All rights

4、 reserved.”Attention is all you need”paper,2017Vaswani et al.,2017,https:/arxiv.org/abs/1706.03762 Massive parallelization Sequential layers were replaced with self-attention mechanisms,for much more parallelism.RNN vs TransformerTransformer Architecture 2025,Amazon Web Services,Inc.or its affiliate

5、s.All rights reserved.Richard Sutton,2019http:/ One can build a custom machine that does very well on a certain task.However,the general purpose approach will eventually outperform the bespoke approach.The Bitter Lesson 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Keep pumping

6、data?Consumed all of text that is available.Now,we re moving onto all the video that it exists.Architecture itself has limitations?Can you name any?Could it plateau?2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Gary Marcus,NY Times Opinion Ed 2025,Amazon Web Services,Inc.or its