Three medical cases and no room
On 27 May 2020 Tarry Singh spoke to Codemotion's online deep learning conference in Italy, carrying three medical case studies and an argument about what neural networks actually are.
Codemotion ran its Deep Learning Tech Conference online on 27 May 2020. No hall, no badge queue, no corridor between talks. What survives from the day is a slide deck, a client case document from an earlier project and a newspaper interview. That is the whole record, and for this year it is the usual shape of one.
Eight days earlier La Repubblica had run a long piece by Jaime D'Alessandro on Elon Musk, Neuralink and the money going into wiring brains to computers. Tarry spoke to him from his house in Groningen. The paper introduced him as a consultant to the World Health Organization and founder of deepkapha.ai and Curae.ai, and told readers where to find him on 27 May. Asked whether artificial neural networks copy the brain, he said they are "an inspiration, nothing more". Then the arithmetic behind the answer: 86 to 100 billion neurons in a human brain, about a tenth of them given over to sight, and roughly a hundred thousand of those simulated so far.
The talk opened away from medicine. Slides on the arms race framing, on states pouring money in while others write law against surveillance, on the wealth that accrued to companies which learned to predict and shape behaviour. Then the correction. AI is already solving real problems at scale, it is not the Manhattan Project, it needs no national programme, and any researcher anywhere can pick up open source tools and work with everyone else.
Three medical cases carried the second half. Breast cancer first, with EPOS Health Management of the GOPA group, on cytology images from Afghanistan, where the project document records the disease as the second leading cause of death among women after maternal mortality. A pool of eight experts ran around thirty models at once on the firm's own hosted servers in Germany, with transfer learning taking accuracy from 82 per cent to about 95. Then diabetic retinopathy in China: a fundus grading system standardised against the British diabetes screening scheme across 70,000 colour photographs, built with researchers in Australia and Singapore, regulator-approved and heading for Tibet. Then teeth, with Charite in Berlin, against 200 million dental X-rays taken in Europe each year and a problem that stays hard at the level of a single tooth.
The partnerships slide listed what was coming next. Kidney and bladder cancer. Endometriosis. And one line reading "Covid19, to be announced". In a deck given in May 2020, that line is the only place the pandemic appears at all.
"La mente collegata al computer, cosa c'e dietro la nuova visione di Elon Musk" by Jaime D'Alessandro. A long feature on Neuralink and brain-computer work, in which Tarry is the sceptical voice, interviewed from Groningen and named as a WHO consultant and founder of deepkapha.ai and Curae.ai. It previews the Codemotion talk by name and date.
Five-page client document, 26 April 2019, on the breast cancer pilot with EPOS Health Management. Source for the eight-person team, thirty parallel models, German hosted servers and the 82 to 95 per cent accuracy figures used in this chapter.
