From the 4th to 12th of July, I was in the city of Porto, Portugal for the 7th edition of the summer school on Vision Understanding and Machine Intelligence (VISUM). The summer school brings together students and researchers with interest in computer vision and machine intelligence. The program was packed with lectures, group hackathon, hands-on project, industry day and social events. In this post, I share notes of the courses, projects, social events and my overall thoughts about the program. The lecture slides are all available on github.
The 6 Courses taken
1. Introduction to Computer Vision and Machine Learning, Jose Costa Pereira
Part I – Computer vision: When compared to image processing computer vision has broader exposure and is a generalisation of image processing. Computer vision can be broadly grouped into low (filtering), mid (e.g segmentation, measuring scene motion) and high-level vision (e.g. scene classification).
Part II – Machine learning (ML): ML views are categorised into 3: the perception, signal processing and communication views. Lecture explains different types of statistical learning (supervised and unsupervised methods).
2. Computer Vision with Deep Learning, Pascal Mettes
The recent breakthroughs in deep learning owe it to large datasets, powerful hardware, open-source software, tricks. Multi-layer perceptrons (MLP) came into the picture to tackle the limitations of perceptrons. The weights and biases are learnt through forward-backward propagation. A forward step maps input to predicted output. The loss step compares the predicted output to the ground truth output. The backward adjusts the parameters to minimise the losses by propagating its gradients. For images, the convolutional neural network is chosen over MLP because of the huge number of parameters images and the local spatial structure that must be preserved. For videos, the temporal element is introduced, hence 3D CNNs and sequential networks are needed.
3. Generative models, Mohamed Elhoseny
The motivation for generative models is that training data are limited. Most real world data have no labels hence unsupervised learning is required. The goal is to use the available training data to generate new samples from the same distribution. Generative models address density estimation which is core problem in unsupervised learning using explicit or implicit models. Generative Adversial Networks (GANs) belong to the implicit density category. The second part of the lecture was on imagination research using GANS, i.e. recognizing unseen categories (Zero shot learning).
4. Optimisation & constraint programming (CP), Deepak Mehta
Constraints on cluster prevents maintenance guy travelling a lot or serving too many customers e.g. capacity constraints, minimal diameter. The first wave of AI was handcrafted knowledge e.g. Gate allocation at the airport considering that some people will need visa. The second wave was statistical models which are not decomposable to a set of constraints. For CP in data mining, the domain experts need to provide constraints for more correct model. Deep learning can be a constraint based model for learning relations and classifier.
5. Visual Approaches to Robotic Control, Dinesh Jayaraman
When making decisions that would be affected by future decisions, sequential decision making is necessary. Also necessary when we know what we want but not how to get it. It can be formulated as a Markov decision process and a control problem (learning a policy). Learning can be done by imitation which comes in several forms: behaviour cloning, inverse reinforcement learning and adversarial imitation
6. Interpretability, Mauricio Reyes
Interpretable ML algorithm is one where the link between the features used by the algorithm and the corresponding prediction is human understandable. It is often used interchangeably with Explanability. They are all about understanding why an AI model produced the result it did and how it reached a decision. Deep learning gives highest prediction accuracy and lowest explanability. Interpretability is necessary for Quality assurance/control, accountability, trust and ethics. With the EU regulations on algorithmic decision making , there is a potential for computer scientists to design algorithms and evaluation frameworks that are explainable and avoid discrimination.
Industry day
On the industry day we entered a challenge to complete all the tasks given by the 8 companies present. Some were quite academic, and some challenging in a fun way like Deloitte’s simulation game. The industry day finished with an insightful debate on the topic “From research papers to real world applications: challenges and difficulties. The group hackathon, sponsored by Bosch, was also launched on the day.
Hackathon
The group competition was on object detection inside a car. Given 10 classes of objects in the training set, the task was to detect known and unknown classes of objects. Also, we were required to correctly classify when the car is empty. For the training set, no unknown objects or empty car examples were provided. We did not have access to the test set. My group (3 of us) came out among the top 5 out of the 16 groups that took part in the competition. We gave a presentation on the final day. The winning team were awarded 3 NVIDIA RTX 2080 DL training kits and best project certificates.
Socials
We had a number of amazing social events during the summer school. On Saturday, 6th July, we had the Pint of science event. It was an occasion for technical topics to be discussed in a very relaxed setting. It was similar to the Databeers I attended during the Machine Learning summer school in Madrid. I always love guided tours around cities I visit. We had one on Sunday through the top attractions of the city of Porto. We went to the São Bento railway station, Clerigos tower, and through the famous Lello library. We ended the tour with a fabulous boat cruise on a traditional “Rabelo”. We traversed the main views of the city and went under the six bridges including Luís. After the tour, we had the social dinner later in the evening. The final social event was the Sunset meetup organised by Data Science Portugal which held at Deloitte office. We had two presenters for the day. João Vinagre spoke on recommender systems while Mohamed Elhoseiny spoke on Imagination Inspired Vision.
Course Credits
The summer school entitles one to 2 ECTS from the Universidade Portucalense (UPT). To qualify for the credits you must take part in the group project. We had to pay a separate fee of 50 Euros for the certificate at the admin office of the university. This is besides the summer school registration fee since this was being issued by the university on a separate account.
Scholarships/funding/cost
There were no scholarships offered; at least none was advertised on their website. But there’s no harm sending them an email to indicate motivation to attend the school and ask if they have any scholarship, need-based or merit-based. I was funded by my research centre to attend. I was told that in this year’s event, the people from academia outnumbered the industry participants unlike previous years. I guess funding may not be very common given its popularity among industry. Registration cost 500 Euros for students.
General remarks
The program had a variety of activities that gave a well-rounded learning experience. The program was nicely structured and the social events were thoughtfully put together. For the future, I think the school needs to improve on the organisation of the hands-on sessions and overall structure of the program. Some of the hands-on tasks felt like being thrown in at the deep end. We had to start from scratch to set up the project environment. Hence, most part of the session was spent installing libraries and dependencies. I also found the program to be tightly packed and intensive above normal capacity. The group hackathon was a demanding task in itself and it was difficult to get ahead without missing out from other activities. The project was launched during the weekend and we needed to have the final results by the next Friday. Most times we had to spend a chunk of the hands-on time working on it.
Overall, I had an awesome experience and a good taste of the city during my stay. The selection of 60 participants from different parts of the globe meant that we had a culturally diverse and sizeable class which one could bond with quickly. Forming a team wasn’t that straightforward for me. I felt a bit of pressure knowing I had just 3 days to form a 3-person group. On the day, I realised that the persons I had in mind had other persons in mind…haha. But I was finally able to join a terrific group (we came out top 5!). The hands-on and hackathon helped with the bonding, but it wasn’t long to find people moving in a trio clique.
