On Continual Learning without forgetting via on-policy self-distillation for Semiconductor Circuit design

​Classical Supervised Fine-Tuning (SFT) with expert demonstrations is typically off-policy and comes with the risk of catastrophic forgetting.  Literature suggests on-policy learning as a potential approach to mitigate this effect.

​Recent body of work on on-policy self-distillation wherein a single model does dual roles:

• ​“Student” model for generating rollouts

• ​"Teacher" model is constructed using privileged information, environmental feedback, or in-context learning.

​This exploits the model's capacity for hindsight rationalization, enabling the teacher to generate dense learning signals for samples produced by the "student". However, this comes with the risk of leakage of privileged information which has to be mitigated.

​The talk explores the ideas and practical challenges of implementation as applied to Semiconductor Circuit Design.

​About Speaker:

​Krishna Sankar M is an engineer with 20+ years of experience in signal processing for wireless and biomedical systems built on integrated SoCs.

​Moved into AI around 8 years ago and is CTO at Maieutic Semiconductors, developing self-improving agentic AI solutions for analog circuit design.

​Maintains a blog on signal processing/AI: https://dsplog.com

​Pre-read

1. ​Self-Distillation Enables Continual Learning

2. ​Reinforcement Learning via Self-Distillation

3. ​Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models

​Join online: meet.google.com/wkp-cjos-kpq

​Looking forward to seeing you!