Six days at UT Dallas, ending in a chatbot hack

The deck is titled "Real world AI Project, UTD Dallas 2020" and lays the week out as six days. Day one was AI fundamentals, machine learning and deep learning theory with a gentle introduction to computer vision and convolutional networks. Day two took CNNs further. Day three was generative adversarial networks and the start of natural language processing. Day four was advanced NLP with time series introduced alongside it. Day five was time series forecasting, ARIMA and recurrent networks. Day six was the hack. Every day carried a theory block and a coding session, in that order.

The photographs run from 5 to 11 January 2020, and they say something the deck does not. This was a big room. Long curved benches four and five rows deep, a laptop open at almost every seat, and the front screen wide enough to carry a Jupyter notebook at readable size across its whole width. The teaching moved between the lectern and the board, and for large parts of the week the screen showed live code rather than slides.

The slides that survive have a sense of humour about the gap between a course and a job. One is headed "Real datasets are a mess", and makes the point that real data does not arrive prepackaged like a tutorial set, that each batch may need its own pipeline, and that this is what you meet on the first day of the job rather than the last day of the course. Another sets classification against semantic segmentation, object detection and instance segmentation on the same photograph, under the heading "Classification's complicated cousins". A third, on putting a model in front of a person, is headed "Production: give your model a good interface", and lists Streamlit, Dash and Plotly.

Day six was a chatbot. The reading was Weston and colleagues on prerequisite toy tasks for question answering, and Sukhbaatar and colleagues on end-to-end memory networks, and the working files are the bAbI question and answer sets with a challenge notebook and its solution. The deck's own instruction for the day is short: create a hack solution for the problem set at the beginning of the week, then winners announced and a certificate for everyone.

That last part is why this chapter has ninety-six photographs rather than eight. Fifty-nine of them were taken on 10 January, and almost all of those are the same frame repeated: one participant at a time at the front of the hall, holding a certificate, next to Tarry Singh. Different person every time. It is the least varied set of photographs in the whole archive and the most worth keeping.

Two hundred and forty photographs survive from the week. Ninety-six are published here: burst repeats were collapsed, and a handful of restaurant frames left out. The deck sets out six days but does not date them, so the range given is the span of the photography.