Optimizing Wireless Inference with FREDI

The FREDI framework manages how multiple wireless devices offload convolutional neural network inference to an edge server. To resolve the tension between inference confidence, network latency, and power usage, FREDI uses a utility function to assign a scalar value to every offloading decision. This function uses a weighted sum, where gains in classification accuracy are balanced against the costs of transmission energy and latency. The goal is proportional fairness, ensuring the framework prioritizes overall system utility over the throughput of any single device.

The Mechanism of Dual-Threshold Inference

Most edge-based inference systems use a single confidence threshold to determine if a result is reliable enough to stop processing. FREDI uses two distinct thresholds, T1 and T2, to manage data transmission. The device performs an initial layer-wise computation to generate a confidence score. If the score is below T1, the data is discarded as non-critical. If the score falls between T1 and T2, the device offloads the intermediate feature map to the edge server. The server processes the data using the deeper layers of the neural network to refine the classification. If the score exceeds T2, the device accepts the result locally. This multi-level approach ensures that only high-value feature maps occupy the wireless channel, reducing unnecessary transmission of redundant data.

Solving the Resource Allocation Problem

FREDI models the resource allocation subproblem as a mixed-integer exponential-cone program. The framework identifies link feasibility by checking the signal-to-interference-plus-noise ratio (SINR). If the SINR is too low to support the minimum throughput required for reliable inference, the link is dropped early, narrowing the search space for the solver.

To handle the coupling between power and bandwidth, which creates a non-convex space, the framework uses an exact projection technique. This transformation relaxes the non-convex constraints by mapping them into a convexified space, effectively linearizing the relationship between transmission power and achievable bandwidth. By transforming the problem in this way, the solver avoids the oscillation of free variables and reduces the complex non-convex objective into a tractable form. This allows the system to compute optimal power allocations for up to 144 user devices in under 0.1 seconds, ensuring global fairness across the network.

Constraints of Centralized Coordination

FREDI treats communication as the primary bottleneck. In scenarios where bandwidth is abundant but compute is scarce, the overhead of managing dual thresholds may exceed the gains of offloading. The performance of the framework depends on the frequency of the optimization cycle and the accuracy of the channel state information. Because the solver relies on a static snapshot of link quality, it breaks down if the topology changes faster than the optimization can converge. When the snapshot becomes outdated, the 'globally fair' allocation is no longer grounded in current physical reality, leading to sub-optimal decisions. The primary open question for implementers remains how to adapt this optimization loop to asynchronous environments where link conditions fluctuate faster than the global utility function can be refreshed.