Two days of TensorFlow and PyTorch in Helsinki
A second Helsinki programme for Ari Hovi, five months after the first, taking a room of Finnish engineers from TensorFlow basics to serving models in production.
On 1 and 2 November 2018 deepkapha.ai taught a two-day deep learning programme in Helsinki, hosted by the Finnish training firm Ari Hovi. The course ran 09:00 to 16:00 in a room called Karuselli, advertised on the venue's door screen as "AI development in practice with Deep Learning and TensorFlow and Pytorch". The programme document was prepared on 28 October, three days before the first session. The audience was working engineers and data people, laptops open, writing code from the first morning.
Day one started at the bottom. Neural network architectures, then TensorFlow's own vocabulary: sessions, graphs, variables, placeholders. From there the room built a first network, a linear fit on 2D data for some and a multi-layer perceptron for others, then moved to image classification on MNIST. The teaching notes acknowledged the limit of the format openly. Two days is not a bootcamp. The stated aim was exposure at speed, and every learner left with a take-home exercise.
Day two went further out. Recurrent networks across both TensorFlow and PyTorch, then TensorFlow in production: distributed computing, exporting and serving models, TensorFlow Serving. The class also got two pieces of the lab's own research. ARiA was presented with a researcher joining from Michigan over video to explain the intuition behind it. LAYGo, aimed at removing the learning rate from training altogether, was discussed publicly for the first time in that room.
The shared folders still hold the working material: Keras and CNN notebooks, an RNN MNIST prediction example, Fashion-MNIST experiments, a full CNN walkthrough, and reading that ran from Bishop to Sutskever's thesis. One notebook in the day one folder was written by a participant, Teemu Hirsimäki, working through the hidden layer problem in his own code. The flipchart at the front carried a list of names in red marker: Hinton, LeCun, Schmidhuber, Bengio. On the second afternoon each learner was handed a certificate in front of it.
Cohort size is not recorded anywhere. The room was booked for a maximum of fifty and the photographs show around a dozen people, so no headcount is given. The programme document is marked confidential between deepkapha.ai and Ari Hovi, so only what the photographs and the public course listing already show is described here.
