Deep Architecture Capstone: Design, Defend, and De-Risk a System

Each requirement should connect to one proposed route or control.

The edge line is the one that usually stays vague. It does not have to. MLPerf Tiny fixes four tasks with named models, sizes and quality targets: keyword spotting (DS-CNN, 52.5 KB, 90% top-1), visual wake words (MobileNetV1, 325 KB, 80% top-1), image classification (ResNet, 96 KB, 85% top-1) and anomaly detection (FC-AutoEncoder, 270 KB, 0.85 AUC). It describes itself as “the collaborative effort of more than 50 organizations from industry and academia”. That was 2021.

It fixes how those models are timed, too. MLCommons, the consortium that runs the benchmark, states the rule in one line: “All MLPerf Tiny benchmarks are single stream, meaning they measure the latency of a single inference.”

So batch size one is not a hardship this brief invented. A KB budget is not an arbitrary constraint from a nervous platform team. Both are measurement conditions an industry consortium already scores against, with published quality targets attached. Write your edge requirement as a number in those units, not as an aspiration.

This is a starting hypothesis to test, not the required answer. Five of its six nodes name a route. The bottleneck fusion layer names a budget, and a budget is not a design decision until it is a number.

A published system fixed that number and measured what it bought. In 2021 the Multimodal Bottleneck Transformer sent all cross-modal attention through B = 4 bottleneck tokens. The alternative was letting every token of one modality attend to every token of the other. Nagrani and colleagues measured both: “We note that for early fusion (Lf = 0), bottleneck fusion outperforms vanilla cross-attention by over 2 mAP, with less than half the computational cost.” The resulting model beat the prior AudioSet state of the art by 5.9 mAP, a 12.7% relative improvement. The authors' own institution put the cost side on the Google Research blog as a “50% reduction in FLOPs compared to a vanilla multimodal transformer model”, with “four bottleneck tokens used in our experiments”.

Four tokens is not a magic constant to copy into a cold-storage system. The modalities there were not sensors and thermal images. What transfers is the shape of the claim: a specific bound, a measured accuracy difference against the unbounded alternative, and a measured cost difference. Your fusion node has to produce all three before anyone approves it.

The branches around it are hypotheses of the same kind, and each owes the same evidence. A sensor state-space branch with explicit missingness channels. A multiscale thermal branch. An equipment-graph branch restricted to edges available at the forecast cutoff. A time-aware maintenance-note encoder. Separate risk and abstention heads.

A serious proposal names credible simpler and competing designs.

The first row is not a courtesy. Tree ensembles have been benchmarked head to head against standard and novel deep-learning methods on “a standard set of 45 datasets from varied domains with clear characteristics of tabular data”. The 2022 finding: “tree-based models remain state-of-the-art on medium-sized data (~10K samples) even without accounting for their superior speed”. The authors also released “every point of a 20 000 compute hours hyperparameter search for each learner”. That is what makes the comparison checkable rather than merely quotable.

So there are four rows to defend against. A feature-engineered gradient boosting model over aggregated sensors, graph summaries, note embeddings and image scores. A dense multimodal Transformer that tokenizes every stream into one global sequence. Late-fusion specialists trained per modality and combined through calibrated outputs. And the proposed hybrid.

Each buys something and costs something. Boosting is a strong tabular baseline that deploys easily, and it sees little raw cross-modal interaction. The dense Transformer is flexible and conceptually simple, and expensive at the edge over long sensor history. Late fusion is modular and independently debuggable, and weak at fine-grained fusion. The hybrid matches modality geometry and controls interaction cost, at the price of more interfaces to test and more integration risk.

Every tensor and state should have a declared unit, axis, timestamp, and ownership boundary. That reads like a counsel of perfection. Then look at what already governs monitoring in this exact domain.

A vaccine refrigerator's thermometer is specified to the clause. WHO's performance specification for a 30-day electronic refrigerator temperature logger requires Type 1 accuracy of plus or minus 0.5 degrees Celsius. It sets the sampling contract in one sentence: “The device must measure the storage temperature at intervals not exceeding 10 minutes and log all relevant data at intervals not exceeding one hour.” Its factory-programmed alarms are defined by magnitude and dwell time together, never by a threshold alone. For Type 1: -0.5 C or below for 60 minutes, and +8 C or above for 10 hours.

The band those loggers watch is set elsewhere. CDC's Vaccine Storage and Handling Toolkit requires that “Storage units must maintain the following temperature ranges at all times: Refrigerators between 2 C and 8 C (36 F and 46 F)”, and prescribes a numbered, documented response to any temperature excursion.

Read that as a template, not as background. It specifies a measurement interval, a logging interval, a precision, an alarm defined as magnitude and duration, and the human procedure that follows an excursion. A model that consumes this data and emits an alert is joining an existing contract. Your ledger owes at least the same fields.

The five steps of this section stand. Define one prediction event, with its equipment, forecast origin, horizon and legal evidence cutoff. Write branch inputs down to sensor channels, image resolution, graph snapshot, note window and missingness masks. Declare intermediate states with dimensions, update frequency, cache lifetime and precision. Specify fusion behavior: available modalities, token budget, attention direction, fallback path. Separate the outputs into risk score, horizon, uncertainty, abstention reason and evidence summary.

Imitate the alarm clause hardest. Your alert threshold, like WHO's, should carry a duration as well as a level.

Each experiment should distinguish value from complexity. The list is long because matched-condition re-evaluation reverses architecture decisions far more often than the original papers suggest.

A large set of proposed Transformer architecture modifications has already been re-run in one shared experimental setting, covering most common NLP uses of the Transformer. Narang and colleagues did it in 2021. The abstract reports the outcome plainly: “Surprisingly, we find that most modifications do not meaningfully improve performance.” The variants that did survive were concentrated among those developed in the same codebase, or amounted to relatively minor changes.

That is the standing prior against every branch in the proposed hybrid. A modification that helped in its own paper, in its own codebase, is not evidence that it will help in yours. The ablations below exist to generate that evidence, at matched latency and matched conditions. A result showing no meaningful improvement is a successful ablation, not a failed one.

The final submission should let another team reproduce the design decision without reading the designer's mind. Steps 1 and 4 — matched performance, and thresholds calibrated against cost — are where deployed early-warning systems have actually failed. There is a measured case of exactly that shape.

A hospital forecasting system with a costly false-alarm side and a catastrophic missed-event side is the capstone's trade under a different name. The Epic Sepsis Model is one, widely deployed and proprietary. An external validation across 38,455 hospitalizations put its AUC at 0.63, in JAMA Internal Medicine in 2021. Wong and colleagues report it in a single sentence: “The ESM also did not identify 1709 patients with sepsis (67%) despite generating alerts for an ESM score of 6 or higher for 6971 of all 38 455 hospitalized patients (18%), thus creating a large burden of alert fatigue.”

Both halves of the trade are in that sentence. Sixty-seven per cent of the events missed, while alerting on 18% of all hospitalizations. It was not a single-site anomaly either. Ostermayer and colleagues evaluated the same model across 145,885 emergency-department encounters and reported 14.7% sensitivity, 95.3% specificity, 7.6% positive predictive value and 97.7% negative predictive value within a six-hour window, in JAMIA Open in 2024. Two independent groups, two settings, and a vendor-reported quality claim that survived neither.

So build the package accordingly. Report matched performance by failure horizon, equipment family, site and missingness, and expect the pooled number to flatter you. Measure operating cost the way MLPerf Tiny does: latency of a single inference, energy, and model size against a stated quality target. A throughput average is not that, and batch size one will never see it.

Calibrate alert and abstention thresholds explicitly against inventory loss and technician workload. Use the units the sepsis validations used: how many events missed, and how many alerts issued per hundred units monitored. Then define release gates requiring improvement, stability, failure coverage, rollback and monitoring readiness. Make an independent external evaluation one of them. That is the test both sepsis papers were, and it is the only one the model's own builders cannot grade.

Submit a system diagram, tensor ledger, architecture rationale, baseline matrix, ablation plan, cost profile, risk register, and release recommendation. Mark the assumptions that remain untested.

A defensible conclusion may reject the complex hybrid, if late fusion or gradient boosting meets the quality target more reliably. Architecture literacy includes knowing when not to reach for a clever architecture. There is published cover for that conclusion: tree ensembles that remained state of the art on medium-sized tabular data, and a shared-setting re-evaluation in which most Transformer modifications did not meaningfully improve performance.

The test of the deliverable is whether a stranger can check it. Every design claim should read like the ones this lesson quotes: a named bound, a measured difference, a stated cost, and a source someone else can open.

You should now be able to read an architecture diagram as a set of information routes, bottlenecks, state transitions, and operating costs. You should also recognize where data, objectives, training, and product controls begin.

The next specialized paths apply these primitives inside language, vision, optimization, generative systems, and production workflows. Carry the same demand for matched baselines and failure evidence into each domain. Carry the citations too. The habit that separates a design review from a design opinion is that every number in it came from somewhere a reader can go.