DEV Community: Evan Tung

Attending CVPR 2024: Interesting Talks and What I Learned

Evan Tung — Wed, 07 Aug 2024 20:44:22 +0000

Last week, I attended CVPR 2024, the annual IEEE/CVF Computer Vision and Pattern Recognition conference. More specifically, I attended the AI4Space Workshop and a few tutorials at the conference, as the main conference sessions didn't capture my interest as much. It was a fun and insightful experience, where I got to listen to many interesting talks, check out multiple interesting posters, and overall learn a lot from the conference. I wanted to go over a few papers and tutorials I found interesting from what I attended.

Mitigating Challenges of the Space Environment for Onboard Artificial Intelligence: Design Overview for a recently launched payload

I am interested in the development and deployment of onboard machine learning for satellites, so this paper was particularly insightful for me. This paper explained well the challenges of "managing thermal constraints, ensuring radiation resilience, [and] overcoming the limited communication bandwidth" when running machine learning models onboard satellites, as well as steps taken to mitigate those risks. The actual hardware used for machine learning was the NVIDIA Jetson Nano, an edge AI hardware device that includes a CUDA-compatible GPU.

Temperature control is crucial to computing in space, as extreme cold will shut off the computers, while extreme heat can cause overheating and damage. There is no natural convection in space, so other actions had to be taken to spread the heat. To mitigate this problem, the team added a "carrier frame, a singular aluminium frame designed to closely follow the contours of the module". The frame acts as a heat sink and conducts heat away from the CPU, GPU, and power management chip. Radiation poses another risk, as it can flip bits in the board's memory, leading to data corruption. The team mitigated the radiation risk by storing backups of all files and computing MD5 hashes of the files. If the hash of a file was different, then the file would be replaced with its backup, and if all backups were also corrupted, the satellite would wait to uplink a good version of the file. The final challenge considered was the amount of bandwidth that the satellite could use when downlinking data to ground stations, which slows transfers and limits the number of files that can be sent daily. The mitigation that the team took was to use the "JPEG-XL algorithm for efficient on-orbit compression of image data". This allowed the team to downlink more pictures in a day.

I found the paper particularly insightful because it used a backup and hashing mechanism to ensure the integrity of files on disk rather than only relying on hardware radiation shields to protect from radiation and bit flips. The team also explained their reasoning for choosing the JPEG-XL image compression algorithm well, discussing the options when using JPEG-XL and the tradeoffs compared to other algorithms. Their software approaches to fixing many "hardware" or "physical" problems were very insightful. This paper also won Honourable Mention at the AI4Space workshop, and I believe they did an excellent job!

Transformers for Orbit Determination Anomaly Detection and Classification

This was an interesting application to a common set of space machine learning problems. Transformers are generally known in the world of natural language processing and large language models as models that accept sequential input in the form of tokens and find relationships between tokens. Seeing them applied to the world of space machine learning on time series data was definitely intriguing. The paper itself goes through parts of spacecraft navigation that fit the model of time series data, and applies transformers to orbit determination, anomaly detection, and anomaly classification.

The team tested three transformer models on tokenized time series data. The Measurement Transformer (MT) "builds off the BERT architecture, adapted for time series data by using a time encoding instead of position encoding". The Tracking Pass Transformer (TPT) uses nested Transformer encoders to extract information from short periods of measurements and combine them for anomaly classification. The Vision Transformer (ViT) "divides an image into a set of 2-dimensional patches, then treats this set of image patches as a sequence that can be processed with a Transformer encoder model". All three transformer models were trained and tested on five classes:

Drag – erroneous estimate of the spacecraft’s coefficient of drag.
Gravity – reduction in spherical harmonics degree and order in the estimation filter’s dynamical model.
Maneuver – erroneous finite thrust maneuver direction and magnitude estimate.
Nominal – no dynamical or measurement mismodel present.
SRP – erroneous estimate of the solar radiation pressure scale factor

Overall, all 3 models "achieved over 80% validation accuracy, with the best performing Tracking Pass Transformer models achieving 93% validation accuracy". This signifies the effectiveness of using transformers for time series data in spacecraft navigation.

Personally, I found this paper very interesting because of the novel approach of something that is generally used on NLP being applied in space machine learning on time series data. It will be interesting to see if, in the future, more models or methods in other fields are applied to space machine learning.

Geospatial Computer Vision and Machine Learning for Large-Scale Earth Observation Data

After attending the AI4Space workshop, I then attended this tutorial on geospatial computer vision at CVPR because it was in the same domain as the current Street Level Imagery project that I am contributing to for the American Red Cross. This tutorial mainly covered background knowledge and recent advances in geospatial computer vision on satellite and aerial imagery, and covered standards and methods, such as GeoTIFF for storing geospatial data inside rasterized images, shapefiles for storing geospatial data as well as several current machine learning models being used for geospatial computer vision. There were also discussions about how to participate in the field, such as competitions like SpaceNet.

Overall, while the discussion of geospatial computer vision was very interesting, it didn't align too much with my Street Level Imagery project, as the tutorial focused more on satellite and aerial images and the challenges and advances there, whereas we are focusing more on street level imagery as it is much more accessible and much more frequently updated. For example, while shapefiles are a common geospatial data format, GeoTIFFs are most commonly seen in satellite and aerial imagery. I will be publishing a post about my experience with the Street Level Imagery project and the American Red Cross soon, so be on the lookout for that!

Conclusion: What's Next?

Attending CVPR 2024 was definitely an experience that allowed me to learn more about two new fields that I have become interested in. The AI4Space Workshop allowed me to gain a better understanding of where machine learning is being applied in space exploration and its current challenges. Meanwhile, the geospatial computer vision tutorial gave a great overview of the current datasets, methods, and community surrounding geospatial CV. These sessions not only broadened my knowledge of the field but also inspired new ideas for my ongoing projects. If you are following my blog, you'll hear from me more as I delve deeper into the papers presented at AI4Space, and when I talk about my experience working on the Street Level Imagery project with the American Red Cross and Civic Tech DC.

AWS Certifications: Are They Worth It? Should You Get One?

Evan Tung — Wed, 07 Aug 2024 20:40:07 +0000

If you have spent any amount of time in the tech world, you will know that while there are many standard certifications that you can get if you work in IT, the world of certifications for Software Development is vastly different and much more disorganized, because of the fast pace at which technologies come onto the scene and become deprecated. There are certifications for specific technologies like MongoDB or Snowflake, but not all companies use those technologies, and there is always a risk of learning a legacy technology. AWS Certifications are a little bit different, as they cover a breadth of cloud technologies. I recently got my AWS Certified Developer - Associate, and I'm here to give an explanation about AWS Certifications, my experience getting my AWS cert, and give an analysis on whether or not you should pursue one.

What are AWS Certifications?

AWS Certifications, as you may have guessed by the name, are certifications that demonstrate skills and experience in Amazon Web Services. Most companies now run their software on the cloud, and Amazon Web Services is the original and largest cloud service provider as well as the innovator in this space. As more and more companies migrate to a cloud-first software development approach, they are looking for people who know how to use cloud services to resolve the complexity and choice that the cloud can bring. One of the primary ways that these skills can be demonstrated is with the certifications that the cloud providers offer.

These certifications are split into four levels (Foundational, Associate, Professional, and Specialty) and multiple certification paths. The certification path that I am pursuing is the Software Development Engineer path, hence my choice to pursue the AWS Certified Developer - Associate. They usually cover a broad range of cloud technologies, including VMs and networking (EC2), object storage (S3), databases (RDS, DynamoDB), and serverless functions (Lambda). The main difference that AWS certs have when compared to technology-specific certifications is that they offer the breadth of knowledge about cloud technologies. While certain technologies may come into and fall out of fashion, Amazon Web Services as a whole is here to stay for a long time. Out of the different types of certifications for software development, AWS Certifications are probably the most valuable because of the both the breadth of technologies they cover as well as companies' need for developers skilled with AWS.

How I Prepared For My AWS Cert

Before I dive into how I prepared for my AWS Certified Developer - Associate exam, you should know that I am not a complete newbie to software development or Amazon Web Services. I have past experience as a Software Dev Engineer, and have used multiple Amazon Web Services since I work there after all. As a result of this, even though the Software Development Engineer certification path starts off with the AWS Certified Cloud Practitioner, I decided to skip the Foundational-level cert and go straight for the AWS Certified Developer - Associate. The certification paths document shows that the AWS Certified Cloud Practitioner is optional for IT/cloud professionals, so if you have already worked with AWS before, you should probably skip it too and go straight to the Associate-level cert for your path.

The entire process of my preparation for the certification exam took around 3-4 months, although I wasn't consistent in my studying for the exam. I decided that I was going to study for my certification in October 2023, and finally took the exam and passed in January 2024. I had experience with most of the popular services from my full-time job and my previous summer internship in 2022. As an Amazonian, I could get 50% off the exam fee as well as free access to the major AWS courses from AWS Skill Builder, A Cloud Guru, Cloud Academy, and more. I started studying the A Cloud Guru course for the cert in October, and went through the videos in my spare time. I had decided not to follow the labs or demos, because the exam itself was just multiple choice questions and at work we use Infrastructure-as-Code with the AWS Cloud Development Kit (CDK) instead. It took me 3 months on and off to finish watching the videos and taking the quizzes, and I would say that the content of the course was spot-on compared to the exam questions. I spent January taking any practice exams I could find, and got an average of 85% correct. I would say that anything above 80% correct is usually a passing score. I finally registered and took the exam in late January, and passed the exam and got the AWS Certified Developer - Associate.

Although I didn't realize it at the time, studying for the AWS certification has actually helped me in my day-to-day job. I began recognizing the best service for solving a problem, and the best practices for using certain services (structuring tables and indexes in DynamoDB, for example). I was able to ask better questions about design decisions and challenge the choice to use one service over another. The resume impact of the certification has not been as big for me because I already work here, but I definitely have more context and knowledge in my day-to-day job.

Should You Get AWS Certified?

This is a hard question to answer, and the short answer is "it depends". But there are a few criteria that you can use to judge yourself on if you should pursue an AWS certification.

Do you already work with cloud services? This is the obvious question. If you are a DevOps Engineer, Site Reliability Engineer, or Software Engineer who also works in operations, then pursuing an AWS Certification is probably a good choice for you. It will help you understand design decisions and choose the right service for the job, like it did for me. It is also a decent resume booster, as companies nowadays are looking for cloud experience and certifications are the most valuable demonstrator of cloud experience.

Are you Early Career? If so, I would also recommend that you pursue an AWS certification. As with the case above, having an AWS certification can be a great resume booster and help you stand out in a sea of junior engineers. As a junior, you need every edge over the competition you can get, and an AWS Certification can be one in your arsenal. It will also help you onboard faster at a company that uses AWS, as you will already know how to use the tools that they are working with.

Is there a certification path for you? Lastly, if there is a certification path that matches your role or career path, then you can maybe consider pursuing the certifications in that path. If you think you would benefit from gaining certifications pertaining to your role or career path, that is something you can consider.

I hope this post helps those who are interested in AWS Certifications and are wondering about whether or not it is worth it to pursue one. My experience with certification may be a little different than others', but I hope that it can still help you decide. If you have any questions or want to relay your own experience with AWS Certifications, please feel free to comment here.