Bedrock support for Functor models
I have a new model architecture called Functor models, the idea applies to many model types in the micro and SLM spaces. A functor model learns by modifications to its function rather than parameters. As a consequence, functor models learn one-shot by nature. These modifications I call deltas are functions themselves so the new model function on learning is f' = f + delta where "+" (circle +) is a direct sum. I still decouple learning from inference in the sense that Bedrock does from what I read. A functor model can be trained as a stochastic model but the "learn on demand" has been found to be more useful. I am far from being a Bedrock guru, is there a way functor models could be supported? This is a collection of docs on Functor models - https://doi.org/10.5281/zenodo.21466484
- Tags
- Amazon Nova ForgeAmazon SageMaker Model BuildingAmazon Bedrock Custom Model ImportAmazon Bedrock
- Language
- English
asked 8 days ago55 views
- Newest
- Most votes
- Most comments
Based on the available information about Amazon Bedrock's current capabilities, there are a few potential pathways for integrating your Functor models:
Custom Model Import: Amazon Bedrock offers a Custom Model Import feature that enables you to bring proprietary models to the platform. You can use your imported custom models alongside existing foundation models through a single, serverless, unified API. This would allow you to access your Functor models on demand without managing underlying infrastructure. This appears to be the most direct path for bringing your novel architecture to Bedrock.
Amazon Bedrock Marketplace: If your Functor models gain broader adoption, the Bedrock Marketplace provides a way to make models available to other users. The marketplace allows discovery, testing, and deployment of models on managed endpoints, with access through unified APIs and native integration with Bedrock tooling like Agents, Knowledge Bases, and Guardrails.
API Integration: All models in Bedrock can use the InvokeModel operation, and many support the Converse operation for more standardized interactions. Your Functor models would need to be compatible with these API patterns.
Considerations for Your Architecture: Given that Functor models learn through function modifications rather than parameter updates, and support one-shot learning, you'll want to consider how this maps to Bedrock's inference-focused model. Bedrock generally separates training from inference, so you'd need to determine whether your 'learn on demand' capability would be exposed as part of the inference API or handled separately.
I'd recommend starting with the Custom Model Import pathway and working with AWS to understand any specific requirements for integrating your unique learning mechanism into Bedrock's infrastructure.
Sources
Amazon Bedrock Model Choice - AWS
Model compatibility - Amazon Bedrock
Using models with Bedrock - Amazon Bedrock
Relevant content
asked a year ago
asked a year ago
Thanks, I really appreciate the response. The LOP (LLM Offloading Pattern) may be a compelling reason to gain support. It uses AgentCore to create knowledge bases consisting of functor LLMs & RAGs. The advantage being a curated, auditable, governable knowledge base layer which LLMs can use in regulatory environments. See https://doi.org/10.5281/zenodo.20056590 Another point is most functor models scale horizontally. This is a live lightsail demo - https://cluster.functormodel.ai/