Intelligence Distribution Network

Abstract

The applications of deep neural networks (DNN) have grown in number in recent years, but there is a need to better support their deployment in the field. This dissertation introduces Intelligence Distribution Network (IDN), a platform specifically designed to support this increasing demand for DNN inference computations. IDN is a fault-tolerant decentralized peer-to-peer (p2p) network that delivers fast, low-cost, scalable execution of DNN models. IDN does this by optimizing DNN inference for this setting, and then distributing the computation among spare computing resources nearby. IDN reduces inference time using BranchyNet, a technique to modify the standard DNN structure with ''exit branches'' at certain locations throughout the network, allowing some samples to exit the computation early and decreasing the inference time.

IDN scales inference to the entire IDN p2p network using Distributed Deep Neural Network (DDNN), a method that allows DNN inference to be distributed among a computing hierarchy consisting of devices, edges, and clouds. IDN achieves fault tolerance and high availability for DNN inference with ParallelNet. ParallelNet can generate multiple DNN model of various sizes, each fitting a device of varying compute capacity. Each device can execute its model independently in parallel. ParallelNet allows DNN inference to be run on more devices independently, improving the fault tolerance and availability of DNN inference. IDN incentivizes the creation of a large number of high quality DNN models that can be shared among users using DaiMoN, a Decentralized Artificial Intelligence Model Network. DaiMoN improves upon today's limited, redundant, and siloed DNN model sharing structures by incentivizing peers to collaborate and share DNN models to improve the accuracy of a given problem in a decentralized manner. Finally, to manage how IDN is used, the ComputeSwap protocol incentivizes peers to participate in the network via an optimistic debt-repayment structure that probabilistically results in repayment based on credits earned for executing inference and debts acquired by using the network. This encourages peers to service inference computations that are needed by other peers by maintaining a balance to not accumulate too much debt or give too little credit.