Pricing the Skin: Cost-Aware Expert Routing for Wireless Neural Implants

A fully implanted brain-computer interface has to choose, every few milliseconds, between two ways of failing. It can think harder on the implant, where heat and power are scarce. Or it can radio more data out through the skin, where every transmitted bit also costs power and heat. Today this choice is fixed at design time. Either a tiny decoder runs inside the body and sends out only a few numbers, or a dumb digitizer streams features to a powerful external computer. We propose making the choice per time step, learned, and priced.

We borrow a tool from large language models. A mixture-of-experts model has several sub-networks ("experts") and a small gating network, the router, that sends each input to only a few of them. Datacenter work on "locality-aware" routing already teaches routers to avoid experts on distant machines, because moving activations across slow links is expensive. We apply the same idea across the most extreme link there is, the skin. Each time bin of neural activity goes to one of four places. It can be dropped. It can be answered by "hold the last estimate" on the external side, which costs no bits at all. It can go to a tiny on-implant expert. Or it can go to a heavy external expert that needs a compressed packet sent through the skin. The router is trained on decoding error plus an explicit transmission price. The result is a variable-rate decoder that spends radio bits only when the external model actually needs new information.

Core Claim

A causal router on the implant, trained end-to-end with a per-expert transmission cost, should give a better trade-off between decoding accuracy and average transmitted bits than three simpler designs: fixed-rate compression, hand-set event-triggered transmission, and any static split between on-implant and off-implant decoding. The gains should be largest on tasks where the gap between a tiny on-implant decoder and a large external decoder is wide. Concretely, the training objective is

$$\mathcal{L} = \mathcal{L}_{\text{dec}} + \lambda \, \mathbb{E}_t \Big[ \sum_{e} g_e(x_t)\, c_e \Big],$$

where $\mathcal{L}_{\text{dec}}$ is the decoding loss (velocity regression or classification), $x_t$ is the binned neural activity at time $t$, $g_e(x_t)$ is the router's probability of sending bin $t$ to expert $e$, $c_e$ is that expert's transmission cost in bits, and $\lambda$ sets the price. The drop, hold-last, and on-implant experts have $c_e=0$ in bits but different on-implant compute costs, which we track separately. Sweeping $\lambda$ traces an accuracy-versus-rate curve. The claim is that this curve dominates the alternatives.

How This Relates to Prior Work

The mechanism comes from datacenter mixture-of-experts systems. LocMoE adds a soft loss that penalizes routing tokens to experts on other nodes. On PanGu-Σ it reports 1.15x to 1.29x throughput gains. Occult hard-limits each token to experts on a few GPUs. C2R groups co-used experts on one GPU so that a token is sent once and used by several experts. Offline integer programming places experts to minimize network hops in Cluster Topology-Driven Placement. All of these assume symmetric, fast fabrics and a roughly balanced load. Our setting differs in three ways. The cost sits almost entirely on one side, the implant. It is energy and heat rather than latency. And routing must be causal and online. We reuse LocMoE's soft-penalty form and C2R's send-once idea: one packet across the skin feeds every external expert. We drop the load-balancing goal entirely.

The zero-bit options come from MoE++. It adds Zero, Copy, and Constant experts that have no parameters and are replicated on every device, so no traffic is needed. MoE++ reports 1.1x to 2.1x expert throughput. We give its Copy expert a physical meaning: it becomes "the external decoder holds its previous latent." That turns a compute shortcut into a transmission shortcut.

Two findings constrain the design. Not All Models Suit Expert Offloading shows three things. Caching works only when consecutive tokens reuse the same experts. Always-on shared experts weaken router decisiveness through a "bypass effect." And local load balancing destroys this consistency. Our always-available on-implant expert is exactly a shared-expert analog, so we test whether it causes the same bypass. Libra argues that balancing is necessary for datacenter efficiency. Here the cost structure deliberately wants imbalance, with most bins held and few transmitted. That is a point where our setting departs from datacenter practice rather than contradicting it.

On the neural side, tinyRSNN / bigRSNN supplies the two endpoints we route between. A 64-neuron spiking network (a network of neurons that communicate in discrete spikes) occupies about 27 KB and reaches mean R² ≈ 0.66 on macaque finger velocity. A 1024-neuron network reaches 0.698. That work also reports 98.36% activation sparsity, which motivates the expectation that many bins carry little new information. The ROLLS decoder shows a 4 mW on-chip decoder in a closed loop, but only for four-class control. Hirata et al. and the Stentrode first-in-human report frame the wireless constraint.

Existing neural mixture-of-experts models such as BrainStack (soft, dense fusion of regional experts for EEG) and NeurIPT (73.5M parameters, server GPUs) use experts for accuracy. They never make experts sparse or place them on separate hardware. The Networked Neuroprosthesis module distributes hardware but uses no learned routing. To our knowledge, nothing in this literature prices the link inside the router.

Research Program

We would build a two-tier simulated system with an explicit cost model. The implant tier holds four components. The first is a causal router, a small spiking or integer network no larger than tinyRSNN. The second is a tiny on-implant expert, a tinyRSNN-class decoder whose output is a few numbers per bin. The third is an encoder that produces a quantized packet of $b$ bits for the external tier. The fourth is the decision logic for drop and hold. The external tier holds the heavy expert, a bigRSNN-class model or a latent dynamics model. It receives packets and, on hold bins, advances its own state with no new input.

Training uses a straight-through relaxation of the discrete routing choice, which lets gradients pass through a hard pick among experts. We sweep $\lambda$ and the packet size $b$. Implant compute is reported in synaptic operations. We do not invent joule-per-bit figures, because the file gives none. Instead, we report results against the ratio of per-bit to per-operation energy as a free parameter. This shows for which hardware regimes routing pays off. Temperature is not measured; power proxies stand in for it, and this limitation is stated plainly.

Three analyses accompany the main curves. First, we measure routing consistency using run lengths of hold decisions and a segment-level analog of Liang et al.'s SRP metric. This tests whether neural state switches, such as the planning-to-movement transitions that NLB '21 targets, break the hold option exactly when accuracy matters. Second, we test robustness to link failure by simulating random and bursty packet loss. A router that already relies on holding should degrade more gracefully than a fixed-rate stream, and we test that directly. Third, we check the bypass effect by comparing a router that can always fall back to the on-implant expert against one that cannot.

Baselines

The naive reference point is to stream every bin's features and decode externally with the heavy model. On the cortical spike train data, that is bigRSNN at mean R² 0.698 as reported in the RSNN paper. This sets the accuracy ceiling at full rate. The opposite pole is on-implant only: tinyRSNN at R² ≈ 0.66, transmitting only decoded velocity. Every method is placed on the same accuracy-versus-bits plane as these two points. Beating them means reaching within a small margin of 0.698 at a rate close to tinyRSNN's output rate.

The strongest competitor on the rate question is a learned fixed-rate bottleneck. The same implant encoder and the same external model are trained end-to-end at a constant $b$ bits per bin. It has an equal parameter budget, equal tuning trials, and the same data splits. It is strong because it captures most of the benefit of learned compression without any routing. The proposal beats it only if, at matched average bits, the routed system is more accurate, or if at matched accuracy it uses fewer bits.

The second strong competitor is event-triggered transmission, also called send-on-delta. It transmits the full packet whenever the change in binned spike counts exceeds a threshold, and holds otherwise. The threshold is tuned on validation data with the same budget as $\lambda$. This baseline isolates whether a learned router beats a hand rule for the drop and hold decisions.

Ablations isolate the key ingredient. The first removes the zero-bit experts, so routing is only between on-implant and remote. The second trains the router with LocMoE-style load balancing but no price. The third replaces the soft price with an Occult-style hard budget. The fourth removes the always-on local expert to probe the bypass effect.

For NLB '21, the file reports co-smoothing of 0.346 bits per spike for AutoLFADS versus 0.187 for GPFA on MC_Maze. It gives no reported behavioral-decoding numbers on Lacuna for MC_Maze, MC_RTT, Area2_Bump, or DMFC_RSG. We would therefore run AutoLFADS-style external models ourselves as the full-rate ceiling for those tasks.

Why not just…?

Why not just stream binned spike counts? They are already cheap. This is the most serious threat. The raw 30 kHz argument does not apply to binned threshold crossings: at around a hundred channels, binned counts are a small stream. If the fixed-rate bottleneck or plain binned streaming matches our curve on NLB-scale recordings, the honest conclusion is that routing matters only at much higher channel counts. We report rate per channel so that the scaling argument is explicit rather than assumed, and we treat parity at current channel counts as a real negative result.

Why not just run tinyRSNN on the implant and send only its output? On finger velocity, the gap between tinyRSNN and bigRSNN (0.66 vs 0.698) is small. For that task, this alternative may be nearly good enough, and we say so. The routed design earns its place only where the gap is wide. Candidates are continuous reaching (MC_RTT) and the cognitive and sensory tasks (DMFC_RSG, Area2_Bump), where NLB '21 suggests simple models struggle. The experiments report the size of that gap per task, and our gain should grow with it. If it does not, the idea fails.

Why not just use event-triggered transmission with a tuned threshold? It is a mixture with two experts and a hand-written router, so it is the natural null hypothesis. If learned routing does not beat a tuned threshold at matched rate, the learned router is not worth its implant compute. The send-on-delta baseline, run with an equal tuning budget, settles this directly.

Decisive Checks

Primary data are the macaque cortical spike train finger-velocity data, plus NLB '21 MC_Maze, MC_RTT, Area2_Bump, and DMFC_RSG. For each, we plot accuracy (R² for behavior; bits per spike for co-smoothing as a secondary measure) against average transmitted bits per second per channel and against implant operations. We compute Pareto dominance over the fixed-rate, send-on-delta, and static-split curves, with confidence intervals across seeds and sessions.

The idea is supported under three conditions. It must dominate the fixed-rate bottleneck on at least MC_RTT and DMFC_RSG. Its advantage must grow with the tinyRSNN-to-heavy-model gap across tasks. And it must degrade more gracefully under bursty packet loss. The idea is refuted if a tuned send-on-delta rule matches it everywhere, or if hold decisions collapse at behavioral transitions. We would detect that collapse as a sharp drop in hold run length together with an accuracy loss localized at movement onset.

As a secondary test, we use SilentSpeech-EEG, where BrainStack reports 41.87% on 24 words. We run regional experts sparsely rather than densely and ask whether C2R-style regional grouping cuts transmitted bits with little loss. This is a proxy for multi-region implants, not an implant result.