Efficient Neuro-Symbolic Policy using In-Memory Computing

Tergel Molom-Ochir, Naman Saxena, Jiwoo Kim, Yiran Chen, Zhangyang Wang, Miroslav Pajic, Hai “Helen” Li
Proceedings of the International Conference on Neuro-symbolic Systems, PMLR 288:384-395, 2025.

Abstract

As artificial intelligence (AI) systems grow in complexity, achieving computationally efficient and interpretable decision-making is crucial. Neuro-Symbolic AI (NeSy) offers a promising framework by integrating symbolic representation with neural learning, but its execution on traditional hardware remains inefficient due to memory bottlenecks and high computational costs. This paper advocates for a paradigm shift in AI acceleration—moving beyond traditional von Neumann architectures toward memory-centric computation, unlocking real-time, scalable, and interpretable decision-making for next-generation AI applications. We envision a future where In-Memory Computing (IMC)-based acceleration fundamentally transforms Neuro-Symbolic policy acceleration by mapping it onto hardware-associative memory, enabling O(1) complexity decision-making with drastically reduced energy consumption and latency. Our preliminary results show that IMC-based symbolic policies achieve up to 100x speedup and six orders of magnitude better energy efficiency than CPU and GPU implementations. Moreover, we discuss how probabilistic symbolic policies can be realized within IMC architectures, enabling AI systems to handle uncertainty while maintaining efficiency.

Cite this Paper

BibTeX
@InProceedings{pmlr-v288-molom-ochir25a, title = {Efficient Neuro-Symbolic Policy using In-Memory Computing}, author = {Molom-Ochir, Tergel and Saxena, Naman and Kim, Jiwoo and Chen, Yiran and Wang, Zhangyang and Pajic, Miroslav and Li, Hai ``Helen''}, booktitle = {Proceedings of the International Conference on Neuro-symbolic Systems}, pages = {384--395}, year = {2025}, editor = {Pappas, George and Ravikumar, Pradeep and Seshia, Sanjit A.}, volume = {288}, series = {Proceedings of Machine Learning Research}, month = {28--30 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v288/main/assets/molom-ochir25a/molom-ochir25a.pdf}, url = {https://proceedings.mlr.press/v288/molom-ochir25a.html}, abstract = {As artificial intelligence (AI) systems grow in complexity, achieving computationally efficient and interpretable decision-making is crucial. Neuro-Symbolic AI (NeSy) offers a promising framework by integrating symbolic representation with neural learning, but its execution on traditional hardware remains inefficient due to memory bottlenecks and high computational costs. This paper advocates for a paradigm shift in AI acceleration—moving beyond traditional von Neumann architectures toward memory-centric computation, unlocking real-time, scalable, and interpretable decision-making for next-generation AI applications. We envision a future where In-Memory Computing (IMC)-based acceleration fundamentally transforms Neuro-Symbolic policy acceleration by mapping it onto hardware-associative memory, enabling O(1) complexity decision-making with drastically reduced energy consumption and latency. Our preliminary results show that IMC-based symbolic policies achieve up to 100x speedup and six orders of magnitude better energy efficiency than CPU and GPU implementations. Moreover, we discuss how probabilistic symbolic policies can be realized within IMC architectures, enabling AI systems to handle uncertainty while maintaining efficiency.} }
Endnote
%0 Conference Paper %T Efficient Neuro-Symbolic Policy using In-Memory Computing %A Tergel Molom-Ochir %A Naman Saxena %A Jiwoo Kim %A Yiran Chen %A Zhangyang Wang %A Miroslav Pajic %A Hai “Helen” Li %B Proceedings of the International Conference on Neuro-symbolic Systems %C Proceedings of Machine Learning Research %D 2025 %E George Pappas %E Pradeep Ravikumar %E Sanjit A. Seshia %F pmlr-v288-molom-ochir25a %I PMLR %P 384--395 %U https://proceedings.mlr.press/v288/molom-ochir25a.html %V 288 %X As artificial intelligence (AI) systems grow in complexity, achieving computationally efficient and interpretable decision-making is crucial. Neuro-Symbolic AI (NeSy) offers a promising framework by integrating symbolic representation with neural learning, but its execution on traditional hardware remains inefficient due to memory bottlenecks and high computational costs. This paper advocates for a paradigm shift in AI acceleration—moving beyond traditional von Neumann architectures toward memory-centric computation, unlocking real-time, scalable, and interpretable decision-making for next-generation AI applications. We envision a future where In-Memory Computing (IMC)-based acceleration fundamentally transforms Neuro-Symbolic policy acceleration by mapping it onto hardware-associative memory, enabling O(1) complexity decision-making with drastically reduced energy consumption and latency. Our preliminary results show that IMC-based symbolic policies achieve up to 100x speedup and six orders of magnitude better energy efficiency than CPU and GPU implementations. Moreover, we discuss how probabilistic symbolic policies can be realized within IMC architectures, enabling AI systems to handle uncertainty while maintaining efficiency.
APA
Molom-Ochir, T., Saxena, N., Kim, J., Chen, Y., Wang, Z., Pajic, M. & Li, H.“.. (2025). Efficient Neuro-Symbolic Policy using In-Memory Computing. Proceedings of the International Conference on Neuro-symbolic Systems, in Proceedings of Machine Learning Research 288:384-395 Available from https://proceedings.mlr.press/v288/molom-ochir25a.html.

Related Material