Technical Roadmap
A Technical overview of Omron's roadmap
Last updated
A Technical overview of Omron's roadmap
Last updated
Currently deployed to bittensor subnet number 2, and to ethereum mainnet, Omron version 1 aims to lay the foundation for verifiable, secure and effective capital networks.
Omron is deployed and running on subnet number 2. Currently, a single LSTM liquid staking duration prediction model is deployed and running fully verifiable inferences leveraging the power of validators and miners within the subnetwork. Validators send queries out to miners, who in turn run verifiable inferences using the validators selected model and input data. Miners then respond with model output and a zero knowledge proof of inference. Validators check to ensure the proof is valid for the model selected and issues a score based upon the performance of the miner, the size of the proof and validity of the proof. Due to the ability to convert a wide variety of artificial intelligence models into verifiable circuits, the scale and scope of Omron is truly limitless.
The two key neurons participating in all bittensor subnetworks are Miners and Validators. Both serve invaluable roles within the Omron subnet, as described below.
Miners
Miners are critical to the Omron subnet, and are responsible for computing outputs and proofs for validator queries. In exchage for honest and performative computations, miners receive increased incentive and thus increased emissions. At a more technical level, miners follow the following run pattern
Request for inference is received via dendrite on the query synapse
The request is converted into a JSON file and sent into a zero knowledge circuit, which produces output along with a proof verifying the inference was run correctly.
Both the proof and accompanying output are provided back to the validator in a JSON object via the synapse.
Validators
Validators are another critical component of the Omron subnetwork. Validators produce requests for inferences, and score responses for miners based on their response time, proof size and proof validity. See the below chart, which depicts scoring for verified and unverified (invalid) inferences. As modeled, scores for miners increase as their proof sizes and response times approach zero. At a more technical level:
Validators produce and provide requests in the form of scaled input features, sending requests to miners via dendrite, targeting the query synapse
Miners respond with output from the circuit, and the validator gathers data about the response including the time it took the miner to respond, the size of the proof provided and whether the proof provided was valid.
Validators score the miner's response based on the following criteria
If the proof provided is not valid, the miner's score will decay as depicted below.
If the proof is valid, the miner's score will recover as depicted below. The rate of recovery and maximum score both adjust based on a performance penalty, which weighs both response time and proof size, placing more weight on response times by default.
On Ethereum, Omron has deployed a single smart contract responsible for accepting deposits of liquid restaking tokens along with wrapped ether. While users have funds locked into the deposit contract, they accrue Omron points. Deposits made to this contract signify a collective desire to improve staking and restaking strategies through the use of verifiable AI. Tokens become eligible to withdrawal at the time of Omron version 2 deployment.
The second version of Omron, version 2, will be deployed to both our Bittensor subnet and ethereum mainnet in the near future. See below overviews for what is expected to change and how these changes will impact users and the broader staking and restaking markets.
The bittensor subnet is improving and iterating daily to foster an ecosystem of healthy competition that produces high quality and performant staking and restaking predictions. Our bittensor subnetwork is core to the deployment of Omron version 2. The network will expand it's support of models, and introduce output based scoring. Effectively, the incentive structure will encourage smaller, faster and higher quality models that are still backed by the trustless security provided by zero knowledge proofs. Models will compete to deliver the best possible staking and restaking optimization strategies across a variety of platforms, most notably across Actively Validated Services deployed within EiganLayer.
Notably, version 2 provides a unique benefit to validators in that validators can craft their own queries to request optimization predictions for their own resources while still enhancing the network as a whole by providing valuable feedback to miners responding to their query. Additionally, validators can accept external queries into the network, delivering more organic traffic to miners and at the same time increasing the performance of the network as a whole.
As aforementioned, two key neurons participate in all bittensor subnetworks; Miners and Validators. Both serve invaluable roles in version 2 of the Omron subnet, as described below.
Miners
Miners remain a critical component of the Omron subnetwork, and compete to provide the most performative verifiable models in terms of speed, efficiency and output quality. In version 2 of the Omron subnet, miners are incentivized not only to produce the smallest proofs in the fastest time, but also to distill their models and circuits down into the highest accuracy model considering performance of the model.
Request for inference is received via dendrite on the query synapse
The miner chooses the model it thinks will perform best, given the staking or restaking strategy optimization task.
The request is converted into a JSON file and sent into the pre-generated zero knowledge circuit for the model selected, which produces output along with a proof verifying the inference was run correctly.
Both the proof, the model hash and accompanying output are provided back to the validator in a JSON object via the synapse.
Validators
Validators remain a core component of the Omron subnet. In version 2, validators share similar responsibilities as in the first version, with added scoring considerations and more potential staking and restaking strategies to choose from.
Validators accept external requests or produce their own requests for inference.
Validators provide requests in the form of scaled input features and target optimization task, sending requests to miners via dendrite, targeting the query synapse
Miners respond with output from the circuit, and the validator gathers data about the response including the time it took the miner to respond, the size of the proof provided and the model's hash. If the validator does not have a local verification key for the model hash provided from the miner, it queries the miner for the corresponding verification key.
Validators score the miner's response based on the following criteria
If the proof provided is not valid, the miner's score will decay as shown in the above version 1 depiction.
If the proof is valid, the miner's score will recover. The rate of recovery and maximum score both adjust based on a performance penalty, which weighs response time, response accuracy and proof size, placing more weight on response times and accuracy by default. The more performant in terms of speed, output accuracy and proof efficiency, the higher the score issued. Validator's will compare output accuracy to a locally run, non-circuitized baseline model.
Outline of Miner and Validator dynamics
Due to the miner's responsibility to provide their own models, they are required to provide verifying keys to validators so they can evaluate the legitimacy of the miner's proofs. The diagram below depicts a constant loop of communication between validators and miners, in which validators poll miners for a list of models they have locally. Based on the list of model hashes received, the validator will compare its local verification keys and determine which if any are missing. For the verification keys not present locally, the validator will query miners which it requires VKs from and download respective verification keys. A VK size limit will be determined on a per-task basis along with a maximum download timeout.
Importantly, the scoring related communication is depicted below. Miners are queried to provide verified output for a specific task, and validators score the miner based on the accuracy of their output to a base model, along with performance and proof verification result metrics.
In version 2 of Omron, we're taking our Ethereum deployment up a level and providing the ability to restake into Renzo via Omron. As aforementioned, users that have deposited their LRT or wrapped ether tokens in version 1 will now have the ability to withdraw those tokens, and migrate their points into Omron version 2.
Users will now be provided the ability to restake into Omron, receive ezETH in exchange for their restake while accruing EiganLayer, Renzo and Omron points for their participation in the network. Renzo is a key player in the liquid restaking space, and through Omron we are providing the ability for protocols such as Renzo to optimize their delegation strategies.
0.02 score maximum
