Summer 2021 Semester Review
The original plan this semester regarding SEDRo is to develop a baseline model to use with the Unity Perception experiments. I also wanted to run a fake agent on all the Unity Perception experiments. Additionally, I also planned to research other early skills that new borns acquire so that I can find and design evaluations for them. I wanted to present a paper called "Probing Physics using Tools from Developmental Psychology" to my labmates.
On the personal front, to improve my knowledge, I planned to a Game Development course from Udemy and a Developmental Psychology course from Edx. I've also wanted to continue reading the Developmental Robotics and Reinforcement Learning books.
During the semester, for SEDRo, I have been researching a baseline model. The model presented in Probing Physics paper seems to fit the bill as the model has been applied to environments similar to SEDRo (IntPhys). The architecture has not been detailed in the main paper, so I have been reading a reference paper called "Generative Temporal Model with External Memory" which explains the architecture in detail. I've also been trying to understand how to apply the VOE method from developmental robotics to artificial agents. For this, I am referring to the paper "Of Bits and Wows: A bayesian Theory of Surprise with Applications to Attention". On the experiments side, I have created two probes for the skills "continuity" and "unchangeableness" in a trail project. Continuity as a physics concept means to understand that an object will trace a trajectory over time and space. Unchangeableness means that objects will retain their properties over time. As planned, I also presented the Probing Physics paper to my team. The paper makes three main contributions. One, a formalization of the VOE method to apply to artificial agents. Two, training and testing datasets designing for each intuitive physics concept. Lastly, introduction of a baseline model that is applied to these datasets. I split my presentation into two parts. the first part focusses on the VOE method, intuitive physics concepts and the database design. My intention on presenting this paper is to collaborate with my teammates to start building a training dataset for at least one of the physics concepts programmatically as the original datasets are huge and contain a lot of variability. The second part of my presentation will be focussed on presented a more detailed view of the formalized VOE technique and the architecture of the baseline model. I am yet to completely understand the architecture of the model.
Apart from SEDRo, I completed 5 sections from the Unity Animation course. I gave preference to this course as it helps me design the experiments. I also completed reading the Developmental Robotics book. I want to keep going back to the book though. It is very resourceful in terms of finding good references for topics. I am re-reading the Reinforcement Learning book by Sutton. Unfortunately, I could not focus building my knowledge in developmental psychology. Writing wise, I wrote a blog about Variational Inference and my experience with learning new concepts.
I also updated my personal website to include all the blogs I have written over the past couple of years. It still needs more work.
Regarding labwork, I feel motivated to work and excel. I took a break for a couple of weeks at the end of the semester. That gave me much needed rejuvenation. My labmates are great as always and are always open to discussion. Dr.Park is supportive and encouraging. I feel it will be helpful to continue working in the lab. I feel connected with the work I am doing and I believe I can make good contribution to SEDRo
My future goal remains the same. I want to work in a research oriented role in the industry. I want to look out for roles that involve simulating human environments and aspects of human behavior. My focus will be at the intersection of developmental psychology + evaluation of artificial agents + Unity platform.
The biggest challenge I faced during the semester was trying to understand the baseline model presented in the Probing Physics paper. However, it is a great opportunity for me to strengthen my basic to get a better understanding of the architecture. On a personal front, I also need to work on being more disciplined about maintaining my journal
Regarding lab operation, I think connecting virtually has provided more flexibility for me and gives me more time to work. I think it would be great if we communicated more amongst ourselves (labmates) and kept each other update to date on our work. More brainstorming sessions between us would be helpful.
As feedback for Dr.Park, since a few of us are interested in industry oriented careers, it would be great if you can point us to relevant internships opportunities that you come across.
I am trying to follow the lab culture to the best of my abilities. I collaborate with my lab mates whenever possible. I try to make my work reproducible. I am earnest in my work. I am working on making my presentations better so that they are easy to understand. I have the drive to learn and excel.
For the upcoming semester, I plan to understand and replicated the baseline model presented in Probing Physics paper. I plan to apply this model to the IntPhys dataset. I also need to get a deeper understanding of the formalized VOE technique. Regarding evaluation design, I plan to collaborate with Rubel to programmatically develop a training dataset for 1 physics concept in SEDRo.
Other than SEDRo, I will be on the look out for a developmental psychology course that is tailored to development in the first 1 year of human life. In terms of courses, I want to complete the Animation course from udemy.
I feel I am in a position to organize my time better. I will make efforts to organize my research time better, especially since I am not taking classes anymore and I have the flexibility of work from home.