Training AI with ICP

General
1

Hello,

I have question, could I train an AI with ICP computing power? eg with a canister or so?

regards

2

Welcome to our forum Solipsos!

Currently, training ML models on IC would be both challenging and inefficient.

  • challenging: you’d need to run the training inside WASM runtime, while it is possible, my guess the support for that may still be immature, however it’s hard for me to be sure of that, cause I didn’t play with it. (another example)
  • inefficient:
    1. IC offers no first-class support for parallel computation - you’d need to code that from scratch.
    2. there is no way to attach TPU or GPU to IC currently, so all training would have to happen on CPU, which cannot scale too far - you’d maybe be able to train some hello world project (e.g. Iris flower classification), but I’d be surprised if you’d be able to go much further than that.

similar discussion: AI and machine learning on the IC?

3 Likes
3

For those interested in training ML models in the IC, I would like to introduce Motokolearn, a Motoko package meant to facilitate on-chain training and inference of machine learning models where having a GPU is not a requirement: Introduction of MotokoLearn V0.1: on-chain training and inference of machine learning models

5

Hi Solipsos,

Training an AI model with ICP (Internet Computer Protocol) computing power, such as using canisters, would be quite challenging and inefficient at this stage. Running machine learning training inside a WASM (WebAssembly) runtime could be possible, but the support for it is still likely to be limited and not optimized for the heavy computational needs of AI. For now, using traditional setups with GPUs or TPUs would be more effective for ML tasks.

ProX PC