Executive Meetings, Hackathons, and Indulging in Transformers 🤖
Vinaya Sharma | January Newsletter 🥂❄️
Hey! If you’re new here, I’m Vinaya and welcome to Bits and Bytes! I’m a 17 y/o emerging technology enthusiast hoping to change the world! 🚀
In this monthly newsletter, I share bits and bytes of my learnings to help you grow from my key insights! 🌱 Stay tuned for bits of self-improvement techniques, and bytes of groundbreaking tech innovations.
January has been all about creativity, exploration, and building. Check it out!👇
Presenting to Walmart Executives 🛒👩💼
Last month I shared an in-depth review of our submission for the Walmart Challenge, comprising of an improved app, an interactive recipe shelf, and a queue management software. We heard back from the judges in early January that we were selected as 1 of the 5 global finalists to present to Walmart and we began preparation immediately!
After presenting to the Blue Labs team we were delighted with the feedback we received. They not only appreciated the depth and breadth of our ideas but also provided us with invaluable insight into the process of identifying successful innovations and scaling them effectively. After countless hours put into this project, it’s rewarding to have caught the attention of the Bluelabs team! 🏅
But with each unique opportunity comes its own set of learnings. Here are my top 3:
Never make assumptions.
When working with external stakeholders, particularly companies, it’s important to remember that the same person will not always be following along with you and your work. Always ask questions, understand where everyone is at, and cater discussions, and presentations to address what the listener already knows and wants to further learn about.
This can be as simple as introducing who you are and what you do, to addressing only a specific part of your prepared presentation depending on how much the audience has already caught up on.
Make it real and show your impact.
When working in the real world, make it real. while consulting for Walmart we made sure to get validation and see if our proposed solutions are actually something wanted by the audience, by reaching out to industry leaders and end users.
Although we did a lot of this primary research for ourselves, we failed to show this during our executive discussion and rather focused on the problem and our solution on a technical level due to time constraints. Looking back, finding a balance between the two would have been optimal.
When speaking with stakeholders always make sure to show them the stats, and let them in on the research that led you to your decision. Be transparent.
Clear and concise wins the race.
Your audience should understand what you’re talking about within the first 30 seconds. Projects can get complicated. Projects have a lot going on. But when you’re sharing your findings or making a presentation, always give a summary first - implant the idea, and then dive in. The simpler the better.
Even better if you support ideas with images, diagrams, and prototypes. I can’t stress this enough and I’m sure I’ve said this countless times in my newsletters before. Pictures speak a thousand words, and no matter how obvious this statement is, many people still fail to implement this in action.
Thank you to Julia Perlotto, Colton Schwenk, and Juan Caride from the Blue Labs team for all your compliments, feedback and advice. Finally huge shoutout to Ian Lockhart for arranging this challenge and for all of the insightful post-presentation conversations.
Building Out a Redesigned Refugee Management System at The TKS Focus Hackathon 👩💻
This month I got the opportunity to participate in the TKS Global Focus Hackathon, and I must say my team Tanvi Sheth, Reeya Pandya, Vani Grover, and Praveena Chenthooran deserve massive shoutouts for the project we put together.
Given 24 hours we diligently put together a deck, pitch and multiple prototypes for EmpowerRefuge, an AI and Blockchain-based reimagined Refuge crisis management system.
The problem and solution we focused on:
With 26 million refugees worldwide and only 1% able to rebuild their lives, it's clear that our current refugee support system is broken. Faud in the process has resulted in over $380 million in reported losses in Canada alone and the funds are not going to those that truly need them. This is where EmpowerRefuge steps in, introducing a new and improved process to provide refugees with secure digital identities and access to opportunities.
The redesigned process:
EmpowerRefuge begins with refugees entering camps and undergoing an initial screening process using computer vision and the Interpol criminal database to verify their identities.
The refugees are then registered, and fill out applications for assistance which are written to the blockchain through the use of smart contracts. This ensures the security and transparency of the process, reducing fraud and improving the allocation of services and funds to those in need.
Refugees apply for funds through a user-friendly dApp
The computer vision technology used in the screening process analyzes facial biometrics, providing a secure digital identity for refugees. This secure identity opens up new opportunities for them, such as access to education, housing, and employment, which they may not have had access to before.
Although I’m happy with our finished product there is a lot of room for growth. My team and I spent a lot of time refining our solution and even switched our idea 70% into the hackathon. My biggest takeaway from this one is to stay committed. It’s easy to hop around ideas and leave one when it gets difficult. But don’t make this mistake like my team and I did - we wasted a lot of time on the ideation phase, and choosing a problem to solve, which reduced the amount of time we were able to allocate to the presentation of our ideas. So always make sure you understand your task at hand and the most important indicators of your success before begging any project.
Exploring the World of Transformers 🏎
I won't lie to you… ChatGPT might be my newest best friend. From helping me make decisions, to giving me life advice, to explaining confusing topics, ChatGPT and I have been through a lot recently. I think it’s fair to say that ChatGPT doesn't need any introduction; it’s been a couple of months now since its open release to the public and it’s still taking the internet by storm.
After building in AI for the past couple of months I was stunned by ChatGPTs ability to generalize to different use cases and it's better than human-level performance. I had been tempted to look under the hood and understand how ChatGPT was built for a while now, and when I heard it was passing medical exams and the bar, that was it for me - I knew I needed to start a deep dive. 🤿
After a week-long rabbit hole of trying to understand GPT-3 (ChatGPT’s older sibling - and the model ChatGPT stems from), Transformers, and a lot of NLP as a whole, it's kind of funny to admit that I actually didn't even know what GPT-3 stood for at the beginning (Don’t laugh at me, I'm sure a good portion of you reading this newsletter are in the same boat). Let's start from there then and break down GPT-3 a little.
GPT-3 = Generatively pre-trained transformer 3. It’s a probabilistic model (meaning it can produce many different outputs for the same input) that stems mostly from the Transformer architecture described in the “Attention is all you need” paper from 2017.
In the original paper, an encoder-decoder model is proposed for language translation. The encoder first “encodes” the input text and then uses a cross-attention layer to influence the translated text. The decoder's job is to predict the next token. In the case of GPT-3, we just use the decoder as it is just generating new text.
The decoder works by first passing an encoded input sequence through several layers of fully connected neurons to produce a set of attention scores. These attention scores determine the importance of each input sequence element in generating the current output element. The decoder then uses these attention scores to weight the encoded input sequence, and the result is combined with the current decoder state to generate a prediction for the next output element. This process is repeated until the decoder has generated the entire output sequence.
This is a very high-level overview but stay tuned for a youtube video I’ll be putting out soon to get the whole scoop!
Some Takeaways From This Rabbit Hole:
Stop thinking and start doing.
Just by taking that initial step to understand GPT-3 set off an entire chain reaction. I’ve been studying computer vision for the past bit and was intimidated to look at anything NLP, but when I started reading and later building (more on this soon) with NLP, I realized it was all in my head. A lot of AI is interchangeable, and I recognized a lot of the concepts I explored in my AI Art a while ago for my CycleGAN was actually present in Transformers. I actually enjoyed exploring GPT so much that I began exploring BERT, all of its variances, working with different embeddings and more.
Always be curious.
This brings me to my second point. Always be curious. We talk about this a lot at TKS - don't look at rabbit holes as unproductive, instead allow yourself to explore.
“I have no special talents. I am only passionately curious.”
- Albert Einstein.
I spent quite a bit of time understanding the current state of NLP and the math of transformers, and it has given me a new perspective for future projects and how I can improve my current ones. This rabbit hole has even helped me find an appreciation for math - an area of ML I originally feared.
Bytes of Brain Food 🍎
Each month I consume loads of content through articles, podcasts, and videos. Here are my favourites featuring several technological breakthroughs. 💥
Googles MusicLM: MusicLM is an AI system capable of generating music from text descriptions. This cutting-edge AI system has been trained on 280,000 hours of music, making it capable of crafting an immersive musical experience in any genre. Imagine creating a melodic journey with just a series of descriptions, or having the ability to request a soundtrack perfectly tailored to a specific image or scene.
However, it's not all sunshine and roses, as some generated samples may have a distorted quality and vocals still have some room for improvement. Google has no immediate plans for release, due to the complex ethical considerations and potential risks associated with AI-generated music. Nevertheless, it's exciting to imagine the possibilities and developments to come in this field!
Microsofts BioGPT: BioGPT is a cutting-edge and game-changing breakthrough in the field of biomedical natural language processing! This pre-trained language model uses the powerful Transformer architecture and is specifically designed for the biomedical domain. With its ability to generate high-quality, fluent descriptions for biomedical terms, BioGPT opens up a world of possibilities for scientists, researchers, and healthcare professionals alike.
The results of its evaluations on various NLP tasks are astonishing, with it outperforming previous models on most tasks and setting new records for accuracy. BioGPT is not just a model, but a revolution in the way we approach and understand the vast amounts of biomedical literature at our disposal.
Closing Thoughts
As I reflect on this past month, I am filled with a sense of excitement and accomplishment. It has been a journey filled with key learnings and growth, both personally and professionally. Balancing school, exams, and my own personal interests has been a challenge, but it has forced me to prioritize and manage my time more effectively. I am eager to dive back into the grind and continue on this path of growth and self-discovery. I hope you are just as excited as I am for the future, as there are some truly amazing projects and content on the horizon that I can't wait to share with you in the next couple of weeks. Keep an eye out!
Loved working with you, can't wait to see what's next in store :)