AI Projects #4
inzva AI Projects #4 ended on May 30 with an online showcase lasted almost 4 hours, where 22 participants from 6 teams presented our AI community the results of their 3 months’ worth of hard work. This batch we had our own GPU so the participants had the opportunity to run larger-scale models compared to the previous AI Projects.
Starting on March 6, this batch of AI Projects was mostly held online due to the precautions taken against the COVID-19 outbreak. Though we could not meet with each other physically and enjoy welcoming warm weathers of spring and summer in our home, Beykoz Kundura, we held our monthly weeks in virtual environments and the teams had the opportunity to receive thorough feedback from other participating community members.
While 3 teams chose to tackle prestigious challenges such as ACMRecSys, Robust Vision 2020, and NVIDIA’S AI City Challenge, the other teams played with projects of their liking and ended up doing fun demos.
Our online journey was concluded with our biggest online event to date on May 30 where we welcomed up to 90 guests, participants’ peers, academics as well as leading experts working in the industry, who came to offer their insights about the projects.
You can check out the showcase schedule here and presentations here.
Moreover, do not forget to see the presentation videos as well as projects’ GitHub pages below.
#4 PROJECTS
RECOMPY: A FRAMEWORK FOR RECOMMENDATION
The participants of this project created a Python library called ‘Recompy’ which has a set of algorithms to train models. Recompy only depends on NumPy and it has all functions in order to train and test models. It has multiple similarity measures to calculate user similarities and it has special metrics to measure the performance of recommender system models.
In addition, the project also included a short demo for guests to play with during the showcase and created a website that showed recommended movies using multiple algorithms in production.
Showcase video: https://youtu.be/O8d6929etWc
GitHub: https://github.com/inzva/recompy
USER ENGAGEMENT PREDICTION (ACM RecSys 2020)
Not everything sailed out smoothly; while 2 participants of this team had to quit due to personal issues arising because of the COVID-19, the remaining 2 proved to be some of the most hardworking AI enthusiasts we ever met!
This team downloaded the dataset provided for the ACM RecSys 2020 Challenge, which is composed of 160 million public tweets and used a deep learning model to predict whether a user will interact with a tweet and if they indeed did, what kind of interaction it would be, more specifically one of the following: Like, Reply, Retweet, and Retweet with Comment.
The members are still working towards securing a place in the challenge’s leaderboard, which makes it impossible for us to post the GitHub page at the moment, nevertheless check out their showcase video where they go into the details about the project below.
Showcase Video: https://youtu.be/B_KfUS1yyRU
CONVERSATIONAL INFORMATION SEEKING (CIS)
This team aimed to do some advance research on conversational search systems. With that in mind, they begin by creating a question answering system, where the goal was to meet the user's needs to answer the questions by looking at the MARCO data which comprises newspaper archives. In order to achieve this, they needed to narrow down large datasets by using search engine algorithms.
The team achieved candidate results from the datasets to answer the given query by the user and defined a similarity model that works with word vectors with these candidate results, then reduced the number of candidates once more. Finally, from these smaller candidate sets, the team tried to estimate the answer to users’ questions by using state-of-art techniques such as BERT.
Showcase Video: https://youtu.be/KnXXlm-jg2s
GitHub: https://github.com/inzva/Conversational-Information-Seeking
ROBUSTNESS OF COMPUTER VISION SYSTEMS(Robust Vision Challenge 2020)
Batuhan Faik Derinbay, Ensar Buğrahan Erdağ, Atahan Özer, Mehmet Ali Özer
The members of this project aim to solve a very current problem with the dataset provided for the Robust Vision Challenge, which is the fact that many dataset-specific state-of-the-art methods lack the ability to perform just as well on different datasets in computer vision. This is due to datasets being limited to particular domains, even though they are gradually updated.
The team aims to develop an object detection network, named RODNet, that is robust and thereby performs well on various datasets; the model’s performance is then benchmarked on four discrete popular datasets.
The deadline for the challenge is the end of July, so the team still continues their progress. Therefore, for now, you can check the presentation video below.
Showcase Video: https://youtu.be/NRtLqMbUAvQ
TEXT CLASSIFICATION USING GRAPH CONVILUTIONAL NETWORKS
This team took a common problem and tried to tackle with an extraordinary approach!
An essential intermediate step for text classification is text representation. As traditional methods represent text with hand-crafted features, such as sparse lexical features, some deep learning models, such as CNN and RNN, aim to learn these representations by capturing semantic and syntactic information in local consecutive words. One drawback of these models is that they may ignore global word co-occurrence in a corpus which carries non-consecutive and long-distance semantics. To overcome this problem the team implemented a graph neural network architecture to preserve global structure information while learning word and document embeddings simultaneously.
Showcase Video: https://youtu.be/7ck1G8AlKzk
TRAFFIC ANOMALY DETECTION (CVPR AI City Challenge)
This team is all about solving a problem that affects our daily lives for social good; they aim to detect anomalies in traffic such as wrong turns, wrong driving direction, lane change errors, and all other anomalies based on video feeds available from multiple cameras at intersections and along highways. The dataset is provided by NVIDIA for Track 4 of the CVPR AI City Challenge and the scope of the project includes background modeling, car detection, perspective modeling, and traffic flow analysis. The members merge all the information coming from these tasks to a decision-making algorithm to obtain final results. Since this is designed to be a social good project, where we will hopefully contact with local municipalities to receive local data, the team continues their work even after the batch was concluded 2 weeks ago.
Showcase Video: https://youtu.be/8XYBPpNarcM
GitHub: https://github.com/inzva/Traffic-Anomaly-Detection
We thank our AI community for all their support and growing together with us by offering their feedback every step of the way under these unexpected circumstances; stay safe and keep your keyboards clean. :)
Our AI Projects will be taking new applications for the fall term starting from July 10 and the new batch will take off around late-August.
inzva is supported by BEV Foundation, an education foundation for the digital native generation which aims to build communities that foster peer-learning and encourage mastery through one-to-one mentorship.
Subscribe to our newsletter here and stay tuned for more hacker-driven activities.