Anchors & AlgorithmsAlsea Bay · Oregon Coast
A passage from the Alsea Bay bar

The pencil and the feed-forward pass

What a fog-bound boat at the mouth of the Alsea and a modern neural network actually have in common.

I. The bar at 0400

Six feet of water and no sky to point to

The bar at the mouth of the Alsea is about six feet deep, and it does not stay where you left it. The shoal shifts with the swell and the river behind it, and on a rising tide the sand fills in faster than the chart can correct.1 Cross it on a clear morning and you have a sandstone cliff on your starboard hand — one hundred feet of it, covered with trees, rising on the south side of the entrance.1 Cross it in the kind of fog this coast keeps in reserve and the cliff is gone. The sky is gone. The horizon is gone. You have your compass, you have the wheel under your hands, and you have a pencil.

The pencil is the part that surprises people. The compass tells you where you are pointed; the pencil tells you where you must be by now.

II. The old discipline

Heading, speed, time, lead line

The discipline the pencil belongs to is called dead reckoning. The Smithsonian's navigation collection lists the four tools plainly: a compass for heading, a sandglass for time, a lead line for water under the keel, and a pair of dividers to step distance off the chart.2 With those, a mariner could keep track of his ship by setting one number against another — so many minutes on this heading at this speed, marked off in straightedge segments against the chart's grid.

Columbus did it this way.2 So did every captain in The Mariner's Mirrour, the 16th-century pilot book the same museum cites as the moment the method was first set down for anyone who could read.2 So did the lifeboat crews the United States Life-Saving Service stationed at Yaquina Bay in 1896, twenty miles north of the Alsea, when fog dropped on the central Oregon coast and there was a boat in trouble somewhere off Cape Perpetua.3

Here is the new word for the engineer reading this: state estimate. The whole business of dead reckoning is a state estimate maintained without the help of the world outside the wheelhouse. You set the compass. You time the run. You move the pencil. The state is your best guess at where you are. The world will tell you later whether you were right.

III. The cloud on the chart

Where the pencil mark stops being a point

Now the harder part. The first pencil mark, made just after you cleared the jetties, was as good as your last sighting of land. The second mark is the first mark plus whatever error crept in during the run — a current pushing you sideways, a wind on the cabin top, a compass swung a degree by the wet wool of a sleeve. The third mark is the second mark plus more of the same.

Each mark inherits the error of the one before. After an hour of running blind, the pencil is no longer drawing a point on the chart. It is drawing the center of a cloud — a probability blur that grows wider the longer the fog holds.

Stonewall Bank sits twenty nautical miles west of Newport, just up the coast from the Alsea, and the NOAA buoy moored there as station 46050 reads sea state in 149 meters of water — wave height, peak period, dominant direction, sampled around the clock.4 Those numbers are useful precisely because they cut into the cloud. Each fresh measurement is a chance to shrink the blur. Dead reckoning without measurements is exposure; dead reckoning with measurements is discipline.

The technical name for what is happening on the chart is uncertainty propagation. The pencil is doing it with graphite. The math is doing it with covariance. Same loop.

IV. The filter

Rudolf Kálmán publishes in mechanical engineering, March 1960

In March of 1960 a young Hungarian-American researcher named Rudolf Kálmán published a paper in the Transactions of the ASME — Journal of Basic Engineering, volume 82, pages 35 to 45.5 He titled it A New Approach to Linear Filtering and Prediction Problems. The electrical engineering community had not been interested in his approach; he placed the work with the mechanical engineers instead.6

The paper describes a loop. You begin with a state estimate and a measure of how uncertain that estimate is. You let the system advance one step in time — your model says the state should move from here to there — and your uncertainty grows. Then a measurement arrives from the world. You combine the measurement and the prediction, weighted by how much you trust each one, and you get a new estimate that is tighter than either source alone. You repeat.

This is dead reckoning, written in matrices. The bar pilot's pencil moves the state forward; his sighting of a buoy is the measurement; his quiet recalibration of where the boat must be is the weighted average. He has been running a Kálmán filter, in his head, since before there were Kálmán filters.

Stanley Schmidt at the NASA Ames Research Center understood what Kálmán had built and folded the filter into the Apollo guidance system; from there it went into submarines, cruise missiles, and every modern aerospace stack.6 But the loop was older than the math, and it had been running on the Oregon bar all along.

V. The feed-forward pass

A deep network is a pencil with a thousand hands

A modern neural network is a different shape of the same idea. The model takes an input, passes it through a stack of weighted layers, and produces an output — a classification, a forecast, a steering command. Each layer transforms the previous one. Each transformation costs precision; each carries some uncertainty forward.

A standard network gives you a single best answer. A Bayesian neural network keeps a distribution at every layer — a cloud at every step, not a point — and propagates that cloud through to the end. Recent work on what researchers call Bayesian KalmanNet couples a Bayesian neural network with an explicit Kálmán filter correction so that the network's uncertainty estimate stays honest as new measurements come in.7 The paper notes that this hybrid produces “well-calibrated uncertainties, especially under mismatch” — which is academic language for the moment your model and the world stop agreeing.7

The ML engineer at her terminal at 11 p.m. is doing what the bar pilot is doing. She is maintaining a state, propagating its uncertainty, and waiting for the next measurement to cut the cloud. The variables are larger and the matrices are denser, but the discipline is unchanged: never confuse the estimate with the truth, and never throw the estimate away.

VI. The lamp does not go out

What is shared at the Alsea, at Stonewall Bank, at a terminal in the dark

A boat is working the Alsea bar somewhere this morning. A Coast Guard 47-foot motor lifeboat sits at Station Yaquina Bay, ready to come south if she has to — the station has been on this water in one uniform or another since 1896, when the Life-Saving Service first put a crew at South Beach.3 Out at Stonewall Bank, buoy 46050 is timing the swell in 149 meters of dark green Pacific.4 At a desk somewhere, an engineer is watching her Kálmán residuals fail to converge and trying to decide whether to trust the model or the measurement.

They share a sentence. I cannot see, and I must still know where I am. The sentence has been spoken in fog for centuries. The instruments answering it have changed. The discipline has not.

The pencil keeps writing. The lamp does not go out.