quote: This is the biggest network upgrade Allora has seen. And in one go it can compete with (and outperform) the best chess engines in the world. Initially, this upgrade was designed to extend Allora's capabilities beyond predicting a single number each round (i.e. regression) by also supporting multi-class classification. Concretely, this requires each inference to be a dictionary containing labels, wherein each label has an associated probability that it's the right label. We realized quite quickly that label-probability pairs are generalizable to any type of key-value pairs. As a result, the network now supports any type of dictionary inference with up to 300 key-value pairs. For instance, in financial price prediction, we used to need one "topic" (set of collaborating workers) per asset and timeframe. Predicting the price of 300 assets required 300 such topics. Now, we need a single topic for the top-300 assets, where an inference contains 300 labels (the asset tickers), each carrying the corresponding predicted value. Or a topic can carry 100 labels plus their upper and lower confidence interval bounds. That way, consumers can better interpret Allora's inferences. Or a topic can solve a multi-class classification problem, as this upgrade was originally intended for. The other cool thing is that the network dynamically discovers the label sets. So it's not even necessary to define a fixed label set per topic, but the keys contained in the dictionaries submitted by the inference workers are collected and combined by the network to learn which labels are relevant each epoch. This means the label set can dynamically evolve. A specific example? I love the idea that Allora can now play chess. Each position enables its own unique set of possible moves. With this upgrade, Allora can learn the candidate moves submitted by the inference workers, learn which worker typically performs best in the type of position under consideration, and combine the associated expected game results or engine evaluations to arrive at the best move. Checkmate. | Allora mainnet has upgraded to v0.17. Network predictions can now be returned as labeled outcomes instead of a single scalar. Existing topics remain single-output through migration, and direct emissions integrations move to v10.

quote: This is the biggest network upgrade Allora has seen. And in one go it can compete with (and outperform) the best chess engines in the world. Initially, this upgrade was designed to extend Allora's capabilities beyond predicting a single number each round (i.e. regression) by also supporting multi-class classification. Concretely, this requires each inference to be a dictionary containing labels, wherein each label has an associated probability that it's the right label. We realized quite quickly that label-probability pairs are generalizable to any type of key-value pairs. As a result, the network now supports any type of dictionary inference with up to 300 key-value pairs. For instance, in financial price prediction, we used to need one "topic" (set of collaborating workers) per asset and timeframe. Predicting the price of 300 assets required 300 such topics. Now, we need a single topic for the top-300 assets, where an inference contains 300 labels (the asset tickers), each carrying the corresponding predicted value. Or a topic can carry 100 labels plus their upper and lower confidence interval bounds. That way, consumers can better interpret Allora's inferences. Or a topic can solve a multi-class classification problem, as this upgrade was originally intended for. The other cool thing is that the network dynamically discovers the label sets. So it's not even necessary to define a fixed label set per topic, but the keys contained in the dictionaries submitted by the inference workers are collected and combined by the network to learn which labels are relevant each epoch. This means the label set can dynamically evolve. A specific example? I love the idea that Allora can now play chess. Each position enables its own unique set of possible moves. With this upgrade, Allora can learn the candidate moves submitted by the inference workers, learn which worker typically performs best in the type of position under consideration, and combine the associated expected game results or engine evaluations to arrive at the best move. Checkmate. | Allora mainnet has upgraded to v0.17. Network predictions can now be returned as labeled outcomes instead of a single scalar. Existing topics remain single-output through migration, and direct emissions integrations move to v10.

Source:https://x.com/AlloraNetwork/status/2090440564562878605