Eminent Domain

The trouble with poison oak, Dr. Martin Vale explained to the committee, was largely a misunderstanding.

The plant produced an oily mixture called urushiol. In human skin, its components reacted with proteins, giving the immune system something unfamiliar to attack. Much of the resulting damage was inflicted by the defense. It was an unfortunate chemical encounter rather than evidence that the shrub bore humanity any particular ill will.

“Nevertheless,” said the chairman, scratching beneath his watchband.

“Nevertheless,” agreed Vale.

Vale had spent eleven years studying the problem and had learned when an explanation was over. He advanced to a photograph of a firefighter on a ventilator.

Deer browsed on poison oak. Birds ate the berries. For humans, even getting rid of it could be dangerous: burning a patch carried the oil into the lungs of whoever stood downwind. The firefighter in the photograph had spent three weeks in intensive care. After that summer, Vale had little difficulty getting on the committee’s agenda.

He proposed an engineered plant virus. Its intended hosts were the urushiol-producing species the committee wanted removed: poison oak, poison ivy, and their close relatives. Preliminary tests had killed the target plants without damaging the surrounding vegetation.

“And it attacks only those plants?” asked the chairman.

“Those are the only hosts we’ve identified.”

The chairman underlined only.

“What happens when they’re gone?”

“Without a host, the virus cannot reproduce.”

The chairman put down his pen. A solution that disposed of itself required very little further discussion.

The funding was unanimous.

The field trials went beautifully. Treated plants withered among untouched grasses. Insects emerged unharmed, birds continued feeding, and the soil showed no troublesome residues. Vale’s team produced photographs in which the dead plants were circled, since otherwise it was difficult to appreciate how much better the landscape had become.

Spraying began across the Pacific Northwest that autumn. By winter, the results exceeded every target except complete eradication. Here and there, healthy specimens remained among their dead neighbors.

Vale requested funding to follow the surviving patches through another growing season. The program director returned his application with a question in the margin: were any of these plants sick?

Vale wrote that this was precisely what concerned him.

The director arranged for a grounds crew to remove them and closed the grant.

In March, a field technician brought Vale a cutting from one of the patches awaiting removal. She had carried it in a plastic bag on the passenger seat. She set it on his desk and stopped halfway through explaining where she had found it. When he offered her a chair, she shook her head, then sat down.

Vale called an ambulance. As the paramedics wheeled her out, she held her keys toward him.

“The car’s in a loading bay.”

He took them and told her not to worry.

Then he opened the bag beneath an extraction hood. The leaves were glossy and unmarked.

By evening, he had moved her car into a visitors’ space. The hospital called while he was still holding the keys. The technician had died; did he have a number for her family?

The laboratory found almost none of the familiar urushiol in the sample, which was briefly reassuring, and several unfamiliar compounds, which were not. One escaped readily into the air and damaged lung tissue directly. Previous sensitivity to poison oak made no difference.

For the first time in Vale’s career, the distinction between an allergen and a poison interested everyone.

The surviving plants still carried the virus. They tolerated an infection that had killed their neighbors, and within them the virus had continued changing. Their outward recovery concealed a substantial alteration in their chemistry.

Officials ordered the patches destroyed. Workers wore sealed protective equipment. Each site was fenced, numbered, and scheduled for inspection before the surrounding land could be reopened.

Then a woman died tending tomatoes in a greenhouse outside Portland.

Vale read the report twice before calling the attending physician.

“Any contact with poison oak?”

“Her husband says no.”

“Gardening equipment? Firewood?”

“She grows tomatoes, Doctor.”

Vale sent for a plant. It arrived in a sealed container, healthy and bearing fruit. His staff found a variant of the virus in its leaves, along with the compound that had killed the technician.

He checked the sample number against the shipping receipt, then asked them to run it again.

Further samples arrived from a nursery and an orchard. The altered virus could infect plants well outside its original host range. In those hosts, too, the infection changed what the plant produced. The concentrations varied. The substance did not.

Vale’s team began testing everything they could obtain. A negative result meant that a specimen was uninfected when sampled. It no longer meant that the species was safe.

Shipments were halted after the nursery stock had been delivered. Gardens were destroyed after the compost had been collected. Insects carried the infection across quarantine boundaries while committees negotiated where those boundaries should be.

The maps acquired circles, then shaded regions, then dates after which no reliable reports had been received.

At first, the emergency food program replaced contaminated crops with other crops. Later, its bulletins listed the remaining stocks of food harvested before the spraying began.

The chairman called Vale from an underground facility.

“You said it would run out of hosts.”

“It was supposed to.”

“What about the unaffected species?”

“We haven’t found one that stays unaffected.”

There was a pause. Vale could hear someone speaking in the room with the chairman, asking a question the chairman did not answer.

“Then stop them producing it.”

Vale looked at the sealed growth chambers beyond his desk. There had been a time when that request would have sounded like a grant.

“We’re working on it.”

“How long?”

Vale had an explanation, but he had learned when an explanation was over.

Outside, the municipal sprinklers came on at dusk, as programmed, to conserve water.

By summer, nothing on Earth was allergic to poison oak.

The Need for Machine Learning is Everywhere!

In my work, which is predominantly information technology, the need for Machine Learning is everywhere. And I don’t mean just in the somewhat obvious ways like security or log file analysis. Consider, my work experience goes back to the tail end of the mainframe era. In fact one of my first jobs was networking together what were at the time considered powerful desktop computers as part of a pilot project to replace a mainframe. Since then

Continue reading →

Big Data Spain Presentation

I was invited to give a talk at Big Data Spain in Madrid, Spain on November 18. If you are keeping track, I was in Barcelona at PAPIs.io the previous day, and even worse I hadn’t slept more than 4 hours in the previous 3 days between traveling, presenting, and preparing to present. Still, other than the power going out for a few minutes, the talk went very well.

Thanks to the Big Data Spain team which put together a great conference and an excellent venue.

PAPIs.io 2014 Presentation

I had the opportunity to present a demo of BigML at the first Predictive APIs conference in Barcelona, Spain. I was the very first speaker on the very first day and other than a few problems with the WiFi dropping out on me and the difficulty reading the terminal during the API demo, everything went well. There were some other great talks during the conference as well, the videos for which are also up on papis.io

Data Science Melbourne Meetup

After we launched BigML in Australia, we spent two weeks in Melbourne touring, giving demos, and talking to potential customers. One of the great venues I was invited to was the Data Science Melbourne meetup. They recorded the talk and made it available on youtube. The poor audio quality was my fault – I have to talk with my hands (as you will see) and couldn’t handle the microphone!

A big thanks to Phil Brierley for organizing this event and to GCS Agile, our partner in Australia, for coordinating everything.

Photos

BigML Late Summer 2014 Webinar – Anomaly Detection!

Things at https://bigml.com were busier than ever in 2014, and by mid-year we had already introduced a second machine learning algorithm with our new Anomaly Detector, based on Isolation Forests. Of course there was also a bunch of other new features as well:

– Model Clusters
– Missing Splits
– Anomaly Detector
… And a tease for a few upcoming things:

– Sample Server
– Dynamic Scatterplot
– Projects

The Value of Things (VoT): MassTLC IoT Conference Panel

I was invited to speak on a panel at the MassTLC VoT conference in June. The topic of the panel was “Analyzing data to get actionable intelligence” which was a perfect opportunity to discuss the applicability of Machine Learning to the analysis of the data deluge that the Internet of Things will likely bring.

It was a pleasure to meet the other members of the panel, and the discussion was lively and informative. A video has been promised but hasn’t surfaced yet. I will post it when it is available.

I did come across an article that touched on the panel and includes the following quote:

— excerpt —

Still, similar to the initial buzz around big data, IoT discussions evoke excitement about the wonderful possibilities: new business models! Competitive advantage! Deeper insights! And they often leave out what’s practical, as Poul Petersen, chief infrastructure officer for Corvallis, Oregon-based BigML Inc., noted during a panel discussion on data analytics. Attaching sensors to every grapevine, cargo ship, train car or transformer, “that’s not too hard,” he said. But “how on earth do you get to that last step?”

By last step, Petersen means how you mine sensor data to find what he called those “aha moments,” or correlations between two seemingly unrelated data points. Sensor data is “big in terms of complexity,” giving businesses millions of data points to dig through. “You don’t know at the outset if two things are related,” he said. “Or you may just get it wrong.” It should be noted that BigML is in the business of helping companies make the leap into advanced analytics to find those correlations, but still, Petersen’s perspective was more than an on-message advertisement. As CIOs know from forays into other types of big data, rich insights don’t just fall out of the data—not even for a data scientist.

— excerpt —

BigML Spring 2014 Webinar – Clustering!

This webinar introduced BigML’s K-means clustering algorithm, which was our first unsupervised learning algorithm. I worked hard on this webinar to come up with several easy-to-demonstrate applications of clustering. One of the coolest thing though is our Model Clusters feature which is a great way to “discover” the rules that describe each individual group in the cluster.

– Item Discovery
– Customer Segmentation
– Active Learning

BigML API Webinar Mar 2014

This is the API webinar are promised at the end of the previous webinar when I was talking about “Programmatic ML”. I think it’s really important that even though BigML has such an excellent User Interface, BigML is an API first company; we export the same API at bigml.io that we use internally for our own UI.

I went thru a pretty detailed intro into the API and then showed some examples of using the API:

– Predictive Application
– Python Bindings
– ToyBoost
– BigMLer