After I reached campus I got more information on logistics around the workshop. We were to run the workshop for two weeks, for complete day. We scrambled to make adjustments. We were happy that we had more time in hand. But with time came more expectations and we were under prepared. We decided to add assignments, readings(papers, articles) and possibly a small project by every student to the workshop schedule.
On first day after introduction, Nandeep started with Blockly and concepts of programming in Python. In second half of day we did a session around student's expectations from the workshop. We ended the day with small introduction around data visualization<link to gurman's post on Indian census and observation on spikes on 5 and 10 year age groups>. For assignment we asked everyone to document a good information visualization that they had liked and how it helped improve their understanding of the problem.
For Second day we covered basics of Python programming. I was hosting Jupyter hub for the class on my system, session was hands on and all the students were asked to follow what we were doing, experiment around and ask questions/doubts. It was a slow start, it is hard to introduce abstract concepts of programming and relating them to applications in AI/ML domain. In second half we did a reading of chapter from Society of Mind<add-link-here> followed it with group discussion. We didn't follow up on first day's assignment which we should have done.
On third day we tried to pick pace to get students closer to applications of AI/ML. We started with concepts of Lists, Arrays, slicing of arrays leading up to how an image is represented in Array. By lunch time we were able to walk everyone through the process of cropping on a face in the image using concepts of array slicing. In every photo editing app this is a basic feature but we felt it was a nice journey for the students on making sense of how it is done, behind the scene. In afternoon session we continued with more image filters, what are algorithms behind them. PROBLEM: We had hard time explaining why
imshow by default would show gray images also colored. We finished the day by assigning all students random image processing algorithms from scikit-learn. Task was they would try them and understand how they worked. By this time we started setting up things on student's computer so that they could experiment with things on their own. Nandeep had a busy afternoon setting up development environment on everyone's laptop.
Next day, fourth day, for the first half, students were asked to talk about the algorithm they were assigned, demo it, explain them to other students. Idea was to get everyone comfortable with reading code, understand what is needed for doing more complex tasks like face detection, recognition, object detection etc. In afternoon session we picked up "Audio". Nandeep introduced them to LibROSA. He walked them through on playing a beat<monotone?> on their system. How they could load any audio file and mix them up, create effects etc. At this point some students were still finishing up with third days assignment while others were experimenting with Audio library. Things got fragmented. Meanwhile in parallel we kept answering questions from students, resolve dependencies for their development setup. For assignment we gave every student list of musical instruments and asked them to analyse them, identify their unique features, how to compare these audios to each other.
On fifth day we picked up text and make computer understand the text. We introduced them to concepts like features, classification. We used NLTK library. We showed them how to create simple Naive Bayes text classification. We created a simple dataset, label it, we created a data pipeline to process the data, clean it up, extract feature and "train" the classifier. We were covering things that are fundamentals of Machine learning. For weekend we gave them an assignment on text summarizing. We gave them pointers on existing library and how they work. There are different algorithms. Task was to experiment with these algorithm, what were their limitations. Can they think of something that could improve them? Can they try to implement their ideas.
WEEK 1 ENDED HERE
We were not keen on mandatory student attendance and participation. This open structure didn't give us good control. Students would be discussing things, sharing their ideas, helping each other with debugging. We wanted that to happen but we were not able to balance student collaboration, peer to peer learning and introducing new and more complicated concepts/examples.
Over the weekend I chose a library that could help us introduce basic concepts of computer vision, face detection and face recognition. BUT I didn't factor in how to set it up on Windows system. The library depended on DLib. In morning session we introduced concept of Haar cascade (I wanted to include a reading session around the paper). We showed them a demo of how it worked. In afternoon students were given time to try things themselves, ask questions. Nandeep particularly had a hard time setting up the library on students system. Couple of student followed up on the weekend project. They had fixed a bug in a library to make it work with Hindi.
On Tuesday we introduced them to Speech recognition and explained some core concepts. We setup a demo around Mozilla Deep Search. The web microphone demo doesn't work quite that well in open conversation scenario. There was lot of cross talking and further my accent was not helpful. The example we showed was Web based so we also talked about web application, cloud computing, client-server model. Afternoon was again an open conversation on the topic and students were encouraged to try things by themselves.
On Wednesday we covered different AI/ML components that powers modern smart home devices like Alexa, Google Home, Siri. We broke down what it take for Alexa to tell a joke when asked to. What are onboard systems and the cloud components of such a device. The cycle starts with mics on the device that are always listening for Voice activity detection. Once they get activated they would record audio, stream it to cloud to get text from the speech. Further intent classification is done using NLU, searching for answer and finally we the consumer gets the result. We showed them different libraries, programs, third-party solutions that can be used to create something similar on their own.
We continued the discussion next day on how to run these programs on their own. We stepped away from Jupyter and showed how to run python scripts. Based on earlier lesson around face recognition some students were trying to understand how to detect emotions from a face. This was a nice project. We walked the students on how to search for existing project, paper on the same. We found a well maintained Github project. We followed its README they maintainer already had a trained model. We were able to move quickly and get it working. I felt this was a great exercise. We were able to move quickly and build on top of existing examples. In afternoon we did a reading on
Language and Machines section of this blog:
Let's not forget that what has allowed us to create the simultaneously beloved and hated artificial intelligence systems during the last decade, have been advancements in computing, programming languages and mathematics, all of those branches of thought derived from conscious manipulation of language. Even if they may present scary and dystopic consequences, intelligent artificial systems will very rapidly make the quality of our lives better, just as the invention of the wheel, iron tools, written language, antibiotics, factories, vehicles, airplanes and computers. As technology evolves, our conceptions of languages and their meanings should evolve with it.
On last day we reviewed some of the things we covered. Got feedback from students. We talked about how we have improvised the workshop based on inputs from students and Jignesh. We needed to prepare better. Students wished they were more regular and had more time to learn. I don't think we will ever had enough time, this would always be hard. Some students still wanted us to cover more things. Someone asked us to follow up on handling data and info visualization. We had talked briefly about it on day one. So we resumed with that, walked them through with the exercise fetching raw data, cleaning it, plotting and finding stories hidden in them.