Let AI Draw Your Diagrams: Goodbye Pen, Paper & Whiteboards

PLUS: Can AI help Liverpool FC lift the Premier League Cup?

Howdy fellas!

Lost in translation between a data scientist & a PM, not here!

As we gear up for our third edition of 'The Vision, Debugged,' brace yourselves for another round of mind-blowing insights and cutting-edge revelations! 💥🔍"

Here’s a sneak peek into this week’s edition 👀

  • ✏️ Decoding Eraser.io’s DiagramGPT

  • ⚽ Deepmind & Liverpool FC team up to create an AI assistant for football tactics

  • 🛒 Amazon’s Just Walk Out Shopping is powered by AI (Actually Indians)

  • 📜 FREE AI Cheatsheets for Data Science, Visualization & Neural Networks

Time to jump in!😄

PS: Got thoughts on our content? Share 'em through a quick survey at the end of every edition It helps us see how our product labs, insights & resources are landing, so we can make them even better.

Product Labs🔬: Decoding “Eraser.io’s DiagramGPT”

Spark and Trouble as Product Manager and Data Scientist, often look at and create mountains of flow charts, database entity relation diagrams, architecture diagrams, the list is endless.

If not mentioned yet, both are on the lazier side and seek AI-powered shortcuts. Voila enter DiagramGPT by Eraser. DiagramGPT is like having a personal artist who listens to your textual musings and translates them into visual masterpieces. But instead of wielding a paintbrush, DiagramGPT wields the power of AI.

Product Labs: Decoding Eraser’s DiagramGPT (created by authors)
Tap the pic to get a better view

What’s in it for you?

Diagram GPT has something for everyone, it can create:

  • Flow charts

  • Architecture Diagram

  • Sequence Diagram

  • Cloud Architecture

If you’re very confused about where to start, Diagram GPT also has an excellent set of presets to get that spark of inspiration.

If you thought, that's all not really. How many times have you become Sheldon Cooper playing Pictionary and covered a whiteboard with complex diagrams? Diagram GPT also allows you to upload a picture which is converted into a diagram type of your choice.

How to generate a “Sequence Diagram” automatically from AWS Support Article using DiagramGPT (source: DiagramGPT use case videos)

Not happy with the first cut? No worries!

Of course like all output by generative AI, we are never happy with the first result and it takes a number of prompt tweaks to get it right.

You can either keep editing the prompt directly in Diagram GPT until you’ve got the desired result or else move to Eraser. In Eraser, the diagrams can be edited using Eraser's diagram-as-code syntax which allows editing the structured code file representing the diagram. Not just that you can also add your own elements to the chart such as text boxes, emojis, and new flows.

Simple Prompt gives a beautiful detailed flowchart (source: created by authors)

Hard work doesn’t really make a difference, the output is comparable to the previous, except specific tweaks (source: created by authors)

Image-to-diagram might need a bit of polishing though, not an accurate output (source: created by authors)

What is the tech under the hood?

We know for sure, that Diagram GPT is powered by Open AI’s GPT-4. But beyond that, it’s a reasonable guess what is the workflow. GPT-4 would be converting the natural language input to a structured format, which is then used to generate the diagram. The syntax used for defining the nodes and connections is similar to that used in graph description languages like DOT (used in Graphviz) or Mermaid. However, it does not strictly adhere to any specific language’s syntax. It seems to be a simplified, human-readable format for describing system architecture.

Something along the lines of the “Diagrams” Python library might be turning the cogs to convert the final schema into a flowchart.

What else is in store?

Diagram GPT, also has an API available which you can easily plug into your applications. However, this option is only available for the paid users, which is a bummer ☹️.

Is my data safe?

This is a question we all have in our heads - some of us choose to remain blissfully ignorant while others are more cautious. Taking this worry off their user’s plate, Diagram GPT assures users

  • OpenAI nor Eraser uses your data for training AI models

  • Eraser may analyze usage patterns to enhance its AI features

Key Takeaways

(Screenshot this!)

Identifying an unsolved user problem: While there are many ways to generate images & videos from text there's a notable absence of solutions tailored specifically for creating technical diagrams. DiagramGPT identified this niche scenario promptly and developed a solution for it.

Focus on a simple, clean UX: The UX of DiagramGPT is very simple, making it very straightforward for users to understand the functionality without any ambiguity.

Leveraging existing strengths: The integration with Eraser.io allows utilizing the existing functionalities and also a very simple way to promote the parent product. Users familiar with Eraser.io would be drawn to try the new product as well.

Hot off the Wires 🔥

We're eavesdropping on the smartest minds in research. 🤫 Don't miss out on what they're cooking up! In this section, we dissect some of the juiciest tech research that holds the key to what's next in tech.⚡

If you are a football fan like “Trouble”, we bet you’ve watched the cartoon TV Series “Supa Strikas”. Along with every player's awesome skills, something of crazy fascination was their "Secret Training Compound," overflowing with futuristic AI for training & strategizing. Well, football (or soccer for our US friends) fans, buckle up, because this might be closer to reality than we imagine!

A scene from “Supa Strikas” where the Coach & Professor analyze players & matchups using futuristic AI at their Secret Training Compound (source: YouTube)

DeepMind, the AI powerhouse behind AlphaGo, has teamed up with none other than Liverpool FC (hmm…we’re still confused as to why they chose Liverpool & not Man U) to explore the fascinating intersection of statistical learning, computer vision, and game theory. Their brainchild? - TacticAI, an AI assistant poised to revolutionize football with tactical insights!

A Long Term Vision…

Imagine a super-powered AI analyzing player behavior to optimize team selection, formations, and even set pieces like corners and penalties! That’s what these amazing researchers envision as an Automated Video Assistant Coach (AVAC)! Pretty neat, huh?

Statistical learning traditionally focused on capturing features like player vectors to understand gameplay. TacticAI takes it a step further, combining this with the power of deep reinforcement learning & the strategic brilliance of game theory. Plus, computer vision throws its hat in the ring, enabling player tracking, pose estimation, and real-time scenario reconstruction.

The “triad” of AI research frontiers associated with football analytics (source: Game Plan research paper)

However, the sheer number of players on the field, the vastness of the pitch, and the limited number of goals scored per match compared to the game's duration make the analysis a rather complex problem.

Where does TacticAI Shine?

TacticAI specializes in corner kicks, as they offer an immediate scoring chance, can dramatically shift the game's course, and happen fairly often. TacticAI goes beyond mere predictions (Who will receive the ball? Where will the shot go?) by actually generating new tactics! Think of these as guided suggestions for player positions and movements for more effective corner kicks.

Under the Hood

TacticAI first transforms labeled football data (spatial data from videos, etc.) into graphs that capture player positions & velocities – essentially, everything that matters on the pitch.

A special Graph Neural Network (an algorithm that can learn interactions about the way different nodes in a graph, like social networks or players on the field) then analyzes these graphs, focusing on the crucial player-to-player interactions that make football such a dynamic sport. This architecture is called Graph Attention Network or GAT (introducing ”attention” enables focusing on the most interesting connections in a network).

TacticAI represents corner kick info as a graph with players as nodes & uses GDL to consider all 4 reflections while training (source: TacticAI Paper)

To overcome limited training data, TacticAI leverages Geometric Deep Learning (GDL). This allows the exploitation of inherent patterns and symmetries in the data (you would expect similar results from left & right corners if the players' state on the field is just reflected). TacticAI uses intricate techniques like "frame averaging" and "D2 group convolutions" to make the most of what it has.

Why does this matter?

The results of TacticAI are so impressive that even experts including data analysts from Liverpool struggled to distinguish TacticAI-generated corner scenarios from real ones! More importantly, these experts favored TacticAI's player adjustment recommendations over 90% of the time in both defensive and tracking strategies.

Check out the dips in attacking probabilities of various players before (actual) & after (suggested) TacticAI’s adjustments for the defensive team. (source: TacticAI Paper)

What's exciting is that since TacticAI requires minimal domain knowledge by avoiding intricate feature engineering, it can be readily adapted to other set pieces (penalties in soccer) and even other team sports (penalty corners in hockey or free throws in basketball).

Could TacticAI be the secret weapon behind Liverpool's current Premier League dominance? Can it help them lift the Premier League Cup? Only time will tell, but one thing's for sure: the future of football is getting smarter & more beautiful.

Spark & Trouble are really excited to see how TacticAI evolves and shapes the future of football (& maybe several other sports)!

Key Takeaways

(Screenshot this!)

Interdisciplinary Approach: The integration of statistical learning, game theory, and computer vision can lead to innovative solutions. This encourages the exploration of intersections of different fields.

Intelligent Modeling: Leveraging structure in the problem space can help model a situation well. Graph algorithms are suited for situations involving multiple entities & some ‘information’ being passed between them.

Creatively Overcoming Data Limitations: Exploiting paradigms like symmetry in data, etc. can help combat data scarcity issues by generating highly plausible synthetic data.

Adaptability: TacticAI’s ability to be applied to other set pieces and team sports underscores the importance of designing adaptable and scalable solutions for broader applicability of any research work.

Spark 'n' Trouble Shenanigans 😜

Curious about what your “AI Personality” looks like?
Just answer these 12 questions to know this based on your approach to generative AI:

Share your AI Personalities in the comments!

PS: Spark & Trouble were both identified to be “The Storyteller

Review of the Week🧩

Well, that’s a wrap!
Thanks for reading 😊

See you next week with more mind-blowing tech insights 💻

Until then,
Stay Curious🧠 Stay Awesome🤩

PS: Do catch us on LinkedIn - Sandra & Tezan

Reply

or to participate.