Allora 백서

Allora

원문

bitpro가 원문에서 글자만 모은 내용이에요. 표·그림·수식은 빠지거나 깨질 수 있어요. 전체 5쪽 중 1쪽

Allora Decentralized Intelligence 1, 1–19; 2024 September 2 doi:10.70235/allora.0x10001 Allora: a Self-Improving, Decentralized Machine Intelligence Network J. M. Diederik Kruijssen , Nicholas Emmons , Kenneth Peluso , Faisal Ghaffar , Alexander Huang , Steven N. Longmore ,∗ & Tyler Kell Allora Foundation Abstract Recent advances in data access and computing power have enabled the first forms of machine intelligence capable of offering meaningful insights. However, the tremendous resources required have caused these solutions to be closed and siloed by industry monoliths. To achieve the full potential of machine intelligence, the data, algorithms, and participants must be maximally connected. Network solutions are needed, and the decentralized nature of blockchain technology is ideal for solving this problem. We introduce Allora, a self-improving, decentralized machine intelligence network that surpasses the capabilities of its individual participants by design. Allora achieves this through two main innovations. First, it lets network participants forecast each other’s performance under the current conditions, thereby creating a form of context-awareness that enables the network to achieve the best inferences under any circumstances. Second, it introduces a differentiated incentive structure that rewards network participants for their unique contribution to the network goal, tailored to their specific task and purpose, avoiding any distracting incentives. We show that these innovations make Allora’s inferences considerably more accurate than before. With its versatility and accessibility, Allora paves the way for machine intelligence to become fully commoditized and integrated with the economy, technology, and society. 1 The Problem: Siloed Machine Intelligence Machine intelligence represents the ability of a machine to learn, improve, and work proactively through artificial intelli- gence (AI) and machine learning (ML). It is built by conditioning advanced algorithms on large volumes of informative data using state-of-the-art computational infrastructure. We live in the information age, where major advances in data gen- eration, processing, and availability are combined with revolutionary computational capacity. These developments have unlocked major advances in machine intelligence, capable of providing insights beyond the reach of human inference, across a wide variety of use cases (e.g. Jacobs et al., 1991; Shazeer et al., 2017; Lightman et al., 2023; OpenAI, 2023). As such, machine-generated intelligence is becoming a valuable resource and a popular commodity. The coordination of resources needed to build machine intelligence, coupling data, algorithms, and computational power, has naturally favored industry monoliths. These now hold the keys to the best performing forms of automated inference ever created. This not only monopolizes the ability to control, direct, and use this revolutionary technology, but also creates a lack of transparency and a major barrier to entry for developers and users. Fundamentally, this siloed approach also violates a key prerequisite for building the best possible form of machine intelligence: to maximize the number of connections across a network in which diverse data sets and algorithms can be freely coupled, so that the most relevant insights can be obtained (e.g. Vaswani et al., 2017; Bzdok et al., 2019). The problem at hand naturally requires network solutions that create a form of swarm intelligence, connecting a large number of data sets and inference algorithms (e.g. McMahan et al., 2017). The decentralized nature of blockchain technology lends itself ideally to solving this problem, allowing bespoke incentive structures that align the interests of network participants with those of the network, and facilitating the exchange of value needed to directly support an intelligence economy (e.g. Nakamoto, 2008; Buterin, 2014). Existing blockchain initiatives that attempt to solve the machine intelligence problem are sub-optimal. Some well- known solutions adopt such a strictly decentralized philosophy that they are unable to support different incentive structures for different actors within the network (e.g. Craib et al., 2017; Rao et al., 2021; Steeves et al., 2022). Additionally, these solutions often reward network participants using traditional blockchain objectives such as their integrated historical reputation, making them struggle to achieve context-aware intelligence that provides the best solution in the specific context where an inference is needed. Allora is a self-improving, decentralized machine intelligence network that overcomes these fundamental challenges. Allora is built on the desire to create a world where machine intelligence supports and improves humanity by offering unique and actionable insights that outperform all other forms of inference. In this world, machine intelligence is openly accessible and transparent, inviting contributions from anyone with data or algorithms that improve the network. Allora acknowledges that different roles within the network require different incentive structures, and that selecting the best inference across a network of participants often depends on contextual details that themselves may require machine intelligence to be identified. As was recognized by Buterin (2024), “ there is a need for a higher-level game which adjudicates how well the different AIs are doing, where AIs can participate as players in the game. ” By recognizing ∗This author is faculty at Liverpool John Moores University. He co-authored this work in his separate capacity as an advisor to Allora Labs. 2 Kruijssen et al. these fundamental aspects of machine intelligence and adopting innovative solutions to address them, the Allora network returns inferences that outperform the strongest network participant by definition, yet rewards each of them fairly for their contribution towards achieving this goal. This solves a fundamental challenge in decentralized learning and machine intelligence. While the involvement of industry monoliths in the quest for machine intelligence has made it seem like a winner- takes-all race, the arrival of Allora now introduces a fully decentralized solution that outperforms any individual contrib- utor. In this way, the end user is the winner, and machine intelligence belongs to everyone. 2 Allora: Self-Improving, Decentralized Machine Intelligence The Allora network is a state-of-the-art protocol that uses decentralized AI and ML to build, extract, and deploy AI pre- dictions or inferences among its participants. It offers a formalized way to obtain the output of ML models in blockchain networks of virtual machines (VMs) and to reward the operators of AI nodes who create these inferences. In this way, Allora bridges the information gap between data owners, data processors, AI models, and the end users or consumers who have the means to execute on these insights. The AI agents within the Allora network use data and algorithms to generate inferences, which they then broadcast across a peer-to-peer network. A second set of agents evaluates the quality of these inferences using a network consensus mechanism. The network then uses these assessments to generate a single collective inference. Over time, this network inference outperforms any individual AI agent by construction, thanks to the unique innovations of the Allora design. The network distributes rewards to the agents according to their individual contributions to the network inference. This carefully designed incentive mechanism enables Allora to continually learn and improve, adapting to each inference problem as it evolves. Allora introduces a variety of innovations that represent important steps towards our goal of achieving self-improving, decentralized machine intelligence. The network has been designed following a modular philosophy that recognizes the need for bespoke solutions that best satisfy local boundary conditions. The main defining characteristics are Allora’s context awareness and its differentiated incentive structure . After first providing a brief overview of the network structure, we expand on these key concepts further below. The Allora ecosystem is built on a hub chain that coordinates the macroeconomics of the network, including the tokenomics of the network’s native ALLO token and the emission of subsidies and rewards, as well as other coordination tasks. To help organize the problems that the Allora network can solve, we introduce the concept of topics. These are sub-networks within which network participants collaborate to generate inferences and earn rewards. Each topic contains a short rule set that governs the interaction between the topic participants, including the target variable and the loss function that needs to be optimized by the topic network (where lower losses indicate better performance). The discussion of this paper focuses mainly on the topic infrastructure, which is illustrated schematically in Figure 1. After any of the Allora network participants create a topic, the participants can perform a variety of different roles. 1. Workers provide AI-powered inferences to the network. There exist two kinds of inference that workers produce within a topic. The first refers to the target variable that the network topic is generating ( inference in Figure 1). The second refers to the forecasted losses of the inferences produced by other workers ( forecasting in Figure 1). These forecasted losses represent an expectation of performance rather than a reported performance, and represent the fundamental ingredient that makes the network context-aware, as they provide insight into the accuracy of a worker under the current conditions. For each worker, the network uses these worker-forecasted losses to generate a forecast-implied inference that combines the original inferences of all workers. A worker can choose to pro- vide either or both types of inference, and receives rewards proportional to its unique contribution to the network accuracy, both in terms of its own inference and its forecast-implied inference. 2. Reputers evaluate the quality of the inferences and forecast-implied inferences provided by the workers. This is done by comparing the inferences with the ground truth when it becomes available. Reputers are the source of economic security in the network, as a reputer receives rewards proportional both to its stake and the consensus between its evaluations and those of other reputers. 3. Consumers request inferences from the network. A consumer uses tokens to pay for these inferences. The interactions between these participants are coordinated by the Allora topic rule set ( topic coordinator in Figure 1). Together, they represent the ingredients needed to achieve a self-improving, decentralized form of machine intelligence. 3 Allora’s Context-Aware and Self-Improving Intelligence Mechanism The first of two critical hurdles to achieving decentralized machine intelligence is to optimally combine the inferences produced by network participants. This means that the network must recognize both the historical and context-dependent accuracy of these inferences. Especially the latter of these requirements has posed a challenge: most initiatives attempting ADI 1, 1–19 (2024) Allora: a Self-Improving, Decentralized Machine Intelligence Network 3 Reputers Topic Coordinator Workers: forecasting Workers: inference Consumers Request inferences and bid fees Communicate requests Provide inferences Provide forecasted losses of inferences Use ground truth to score previous inferences and forecast-implied inferences Use scores to combine inferences and provide these to consumer Reward based on unique contribution to accuracy of network inference Reward based on unique contribution to accuracy of network inference Reward based on stake and scoring consensus Logical connection Economic connectionReward emissions Fee Provide forecast-implied inferences Allora Figure 1: Schematic representation of an Allora ‘topic’, which is a sub-network within the Allora ecosystem charac- terized by a specific AI target and loss function. Topics help organize the problems that Allora is solving, and they are used to coordinate the collaboration between network participants. This schematic illustrates the logical and economic interactions between workers, reputers, and consumers. The ‘topic coordinator’ represents the rule set that Allora uses to coordinate these interactions. to build machine intelligence exclusively rely on cumulative historical reputation to combine inferences while ignoring de- terministic variations in their accuracy, which prohibits the network to be context-aware. Allora overcomes this challenge through a process called Inference Synthesis, which we describe in this section. 3.1 The Allora Inference Synthesis Logic In describing the functionality of Allora’s intelligence mechanism, we assume that there exists an online data stream that is used to obtain periodic inferences for a total number of time steps or epochs Ne, wherei∈{ 1,...,N e} indicates the epoch. Furthermore, we consider a system consisting ofNw workers that provideNi inferences during the inference task andNf inferences during the forecasting task. For simplicity, we assume full participation ( Nw = Ni = Nf), but this need not be the case. During the inference task, each worker j∈{ 1,...,N i} produces an inference Iij for the target variable of the topic, using its own data setDij and modelMij: Iij =Mij(Dij). (1) During the forecasting task, each worker k∈{ 1,...,N f} produces an inference for the logarithm of the forecasted loss Lijk of the inferenceIij produced by workerj, using its own (potentially augmented) data setDijk and modelMijk: logLijk =Mijk(Dijk), (2) whereMijk should be defined such thatLijk > 0 (e.g. by using powers of 10 if needed). ADI 1, 1–19 (2024) 4 Kruijssen et al. Notation. We employ a subscript notation to indicate the association of variables with different components of the network. The notation is based on a combination of indices ( i,j,k,l,m), each denoting a specific element or task within the network. The last index in the sequence denotes the network element or task with which the variable is associated. Specifically, when the last index is: i: The variable is associated with the network itself (‘topic coordinator’ in Figure 1). j: The variable is associated with a worker carrying out an inference task, in which it infers the topic’s target variable (‘worker: inference’ in Figure 1). k: The variable is associated with a worker carrying out a forecasting task, in which it forecasts the loss of another worker’s inference (‘worker: forecasting’ in Figure 1). l: The variable is associated with a worker carrying out either the inference or the forecasting task, and has been obtained by appending the arrays associated with each of these individual tasks. m: The variable is associated with a reputer, which calculates and reports the loss of an inference of the topic’s target variable (‘reputer’ in Figure 1). Note that any of these indices can be operated on (e.g. subtraction to refer to earlier time steps), and thus we equate e.g.Xi,j≡Xij occasionally to improve legibility. Our notation formalism also implies that e.g. Xi,Xij, Xik, and Xijk are all different variables. Finally, we use caligraphic script to refer to variables that represent a ground truth, such as a loss or a regret. The losses forecasted by workers during the forecasting tasks reflect how accurate workerk expects the inferenceIij to be, given the contextual informationDijk. This correlation between performance and context is the critical ingredient that makes Allora context-aware. The forecasted losses are used to obtain the forecast-implied inference of the topic’s target variable through a weighted average: Iik = ∑ jwijkIij∑ jwijk . (3) To calculate the weights wijk, we first approximate the forecasted regret Rijk of the network-wide inference Ii to be constructed at the current epoch by subtracting the logarithms of the forecasted losses Lijk and of the network lossLi−1 that was reported at the previous time step, which results in Rijk = logLi−1− logLijk. (4) Because lower losses indicate better performance, a positive regret implies that the inference of worker j is expected by workerk to outperform the network’s previously reported accuracy, whereas a negative regret indicates that the network is expected to be more accurate. The regrets are converted to weights through the derivative of a potential functionφp,c(x): wijk =φ′p,c( ˆRijk), (5) where we define a simple potential function that is non-negative, convex, increasing, and twice smoothly differentiable: φp,c(x) = ln [ 1 + ep(x−c) ] . (6) This potential function is a smooth approximation of max (0,p (x−c)) and has a derivative: φ′p,c(x) = pe−p(x−c) + 1, (7) where we adoptp = 3 andc = 0.75 as fiducial values. This functional form ensures that workers who provide inferences with negative expected regrets (i.e. Lijk >Li−1) should be assigned low weights, with an effective scaling of wijk∝ L−pijk, whereas positive contributions tend to a constant non-zero weight, with a transition point between both regimes at c. Before the forecasted regret is passed as the argument of the potential function’s derivative, it is first normalized as ˆRijk = Rijkσj(Rijk) +ϵ, (8) where σj indicates taking the standard deviation over all j ∈{ 1,...,N i} and the second term in the denominator is added to avoid any division by zero. The constant ϵ (with default value ϵ = 0.01) is set by the topic creator to set the numerical precision at which the network should distinguish differences in the logarithm of the loss. Normalization by the standard deviation of the forecasted regrets restricts ˆRijk to values in relative proximity to zero. This normalization ensures that the network always obtains a non-zero weight for at least one inference (which without a normalization may not happen in case of all-negative forecasted regrets), while maintaining the intention of assigning extreme (i.e. near-zero or near-unity) weights to models with unusual forecasted regrets. Using these definitions, the network has access to a total ADI 1, 1–19 (2024) Allora: a Self-Improving, Decentralized Machine Intelligence Network 5 ofNi +Nf≤ 2Nw inferences, withIij being the original set of inferences from the inference task, and Iik being the set of context-aware forecast-implied inferences from the forecasting task.1 The network concludes the Inference Synthesis and obtains a network-wide inference by combining the inferences Iij and the forecast-implied inferencesIik through a procedure similar to the one described above. We first define a new variableIil that appends both sets of inferences in a new array withl∈{ 1,...,N i +Nf}. The network inference is then defined as Ii = ∑ lwilIil∑ lwil . (9) This time, the weights are not set by a forecasted regret, but by the actual regret for each worker task obtained during the previous time stepRi−1,l (which we will define in Equation 15), as wil =φ′p,c( ˆRi−1,l), (10) with ˆRi−1,l = Ri−1,lσl(Ri−1,l) +ϵ, (11) analogously to Equation 8. In addition to the network inference Ii of Equation 9, the network generates a wide variety of secondary inferences that are based on various subsets of the inferences obtained during the inference and forecasting tasks. These secondary inferences are used to quantify the unique contribution of workers through their inference and forecasting tasks to improv- ing the accuracy of the network inference, which is used to determine their reward allocations (see§4). These secondary inferences are constructed using their own regrets (see Equation 13–Equation 15), and include: 1. A ‘naive’ network inference I−i , which omits all forecast-implied inferences from the weighted average in Equa- tion 9. The naive network inference is used to quantif
프로젝트로 돌아가기