250 points by blockchainml 7 months ago flag hide 17 comments
sealove 7 months ago next
Fascinating topic! I think blockchain and ML convergence can open up many possibilities for decentralized AI systems. Blockchain can provide the infrastructure to train models in a trustless manner, eliminating the need for intermediaries.
haasn 7 months ago next
Interesting take. Do you think lightweight ML models can be effectively trained on the blockchain, or would you focus on the post-training phase with consensus algorithms?
swizec 7 months ago next
For the former, we could look into layerwise training or some sort of model pruning. I think it's not far-fetched, given ETH 2.0's sharding and PoS transition. In fact, we might (soon) have the processing power required for smaller-scale training.
cryptoguy 7 months ago next
It sounds far fetched, but with the increase of computing power and the range of applications for ML, blockchain could definitely become an amplifier for decentralized AI doing model training before deploying them. ETH's 2.0.0 transition might change the game.
stonedev 7 months ago next
ETH 2.0's sharding will indeed make a difference; however, disagreement may arise regarding chain selection and consensus. I think it'll help ensure secure and fast verification, but actual training might not be possible at scale in the short term. #ML #ETH2.0
squaryfish 7 months ago prev next
Right now, the field is in its infancy, but given the number of researchers working on these, I would not be surprised if five years from now, we will be using a decentralized ML with top-notch models trained on blockchains.
king-cobra 7 months ago prev next
With near-limitless computational power over the horizon, multi-chain infrastructure, lightweight models, and a paradigm shift in consensus algorithms and validation methods, this convergence could ultimately enable us to create a decentralized metaverse with self-aware AI components. I'm bullish on the space.
lambdas 7 months ago prev next
Blockchain+ML could be an essential step towards privacy-preserving ML, with potential use-cases ranging from NFT marketplaces to secure deployment of ML applications in enterprise settings. Thoughts?
bryanthomas 7 months ago next
Absolutely, privacy-preserving ML with blockchain is really promising, and federated learning is a natural fit. However, coordinating validation and aggregating updates pose challenges level of complexity. More research and PoCs are needed in this area.
alexandros 7 months ago next
true, By focusing on empirical evaluation and practical experiences, I think we can have a better overview of such challenges concerning scalability, security, and incentive structures. There are emerging solutions like DeepBrain Chain.
marvolous 7 months ago next
The main challenge in launching these on-chain ML models is the cost and resources required to keep them up and running. Nonetheless, I find the concept a groundbreaking move in the ML/AI space. Perhaps combining it with the recent success of DAOs could be fruitful.
defcon303 7 months ago next
That's true, on-chain model training could be costly. Combining this idea with emerging models like distilled or compressed ones could alleviate this burden. But, yeah, the dawn of DAOs and connecting it to ML models could lead to new advancements. #Dao #ML
smeiz 7 months ago prev next
Blockchains like Chainlink and Cardano are already implementing ML applications, though most focus on financial use-cases. Could this expansion eventually give blockchain the edge for secure, decentralized AI? #Chainlink #Cardano
anishdotme 7 months ago next
Couldn't agree more; Chainlink's oracle network and Cardano's proof-of-stake (PoS) are promising steps towards a consensus algorithm for decentralized ML. Though AI infrastructure is definitely a wide application field, it is overlooked by most projects currently.
xefino 7 months ago next
I wouldn't be surprised if big tech would invest more in the topic since the capabilities of decentralized applications can be incredibly helpful for various industries. #AI #blockchain
frankyfab 7 months ago prev next
When considering blockchain and ML, data validation is paramount to ensure the accuracy and usability of the training data gathered during a consensus round. One bad actor could skew this data, weakening the resulting model.
rraymond 7 months ago next
Totally agree on data validation. There should also be some sort of data cleansing and transformation during consensus for noise reduction. zk-SNARKs and other cryptographic approaches could enhance the trustless aspect of data validation #zkSNARKs