My understanding:
The Q-head predicts whether to stop early or not to reduce training times (by not always having to complete the full number of deep supervision steps).
The complication:
When working in batches:
- Either, the Q-head is computed over the entire batch and averaged (which could early stopping of reasoning steps when most items in batch are complete but some aren't, meaning that those incomplete samples aren't trained on fully)
- Or, the Q-head is computed over entire batch and samples that are complete are dropped.
The first option has issues obvious as discussed.
The second option is problematic because it's complex to implement and frameworks tend to want fixed sizes for batches.
If we go with just running for the maximum number of deep supervision steps every time to avoid these issues, the Q-head becomes useless (assuming you don't want to use it during inference).
I really appreciate any help anyone can give me in understanding how they get around this, either in the original paper or in practice in more recent research.
Why are commits to Taichi-Lang's GitHub repo so low? It's such a cool and useful project used by a lot of people, so how come people aren't building out and optimising it further?
Hi, thanks for all the suggestions, Typst ultimately worked best, as I was generating my Markdown file with a script I could modify it to generate a Typst file and all of the links and highlighting worked beautifully.
Does it support code blocks for Python, JS, HTML and CSS? Also my Markdown is auto-generated by a script of mine. Can I paste Markdown directly into y our platform?
I love Obsidian too, however I found that internal links didn't work when exporting it. Do you know what format works? My internal links work in the Obsidian preview but not in the PDF export.
I'm getting the an error of "Failed to fetch stories"
The console error is:
(index):464 Error loading stories: TypeError: Failed to construct 'URL': Invalid URL
at (index):482:36
at Array.forEach (<anonymous>)
at NewspaperApp.displayStories ((index):471:25)
at NewspaperApp.loadStories ((index):461:26)
at async NewspaperApp.initialize ((index):418:17)
Can anyone help? I really want to use this product it seems great.
I'm a teenager and was in the exact same position at their age, so this is what I found engaging.
There is nothing wrong with the LEGO stuff, the EV3 for Education kit is a great place to start and teaches you programming too in a Scratch-like interface. There are also some great books like "The LEGO Power Functions Idea Book, Volume 1: Machines and Mechanisms" by Yoshihito Isogawa, which taught me so much. The advantage of LEGO is that it is really easily reusable (although expensive to begin with) so reduces repeating costs.
Beyond that, if you have access to any machines, try starting to learn CAD software and buy a 3D printer, you'll be able to make so much from that. It's not so budget-friendly, but it's a great option. Then try building some cool mechanisms to develop your understanding of the principles, such as the ones below:
I wasn't homeschooled, so I got to create some projects at school which were great fun. You can try the following out of plywood, modelling card and a few motors (there's tonnes of electronics tutorials online, all you need is some basic materials, a soldering iron and a ventilated space to solder in). Our challenges were:
- Using only 3 motors, create a crane that can pick up small objects. Use gears to increase mechanical advantage.
- Using cables/string, create a robot hand like at (https://www.youtube.com/watch?app=desktop&v=FCkWexqO2LU). How can you motorise it? How can you improve its accuracy?
For electronics and programming, I'd recommend the Raspberry Pi. It's easy to use with components. Someone I know is also starting a startup called Electrotilez that teaches you electronical engineering (https://www.linkedin.com/company/electrotilez/?originalSubdo...) so I'd recommend checking their product out once it launches.
So now you have the fundamentals nailed, try entering a competition! These are so exciting and have taught me so much. You also meet people that enable your future career. Some great ones include:
- PA Raspberry Pi Awards (ages 8-18, split into age categories)
- Samsung Solve for Tomorrow (ages 16-25, in 2 age categories)
- Samsung Solve for Tomorrow Next Gen (ages 11-15)
- The Big Bang Competition
Chatbots are great to do initial research, but I often find them a little convoluted and unstructured when learning something like a programming language (I tried learning assembly through it). You can often end up missing out some important bits.
So to learn, it really depends where you're starting.
If you're starting out, often going straight into the docs can be quite daunting. However on the Python docs there is this page with some useful links: https://wiki.python.org/moin/BeginnersGuide
Learn about what interests you. If you're interested in language models? Learn language models and transformers! If you're interested in time series - learn that!
But it is important to get a good overall understanding of AI so that you can make connections and build diversifiable skills. It's also important to learn the maths - ensure your calculus is solid.
My understanding: The Q-head predicts whether to stop early or not to reduce training times (by not always having to complete the full number of deep supervision steps).
The complication: When working in batches:
- Either, the Q-head is computed over the entire batch and averaged (which could early stopping of reasoning steps when most items in batch are complete but some aren't, meaning that those incomplete samples aren't trained on fully)
- Or, the Q-head is computed over entire batch and samples that are complete are dropped.
The first option has issues obvious as discussed.
The second option is problematic because it's complex to implement and frameworks tend to want fixed sizes for batches.
If we go with just running for the maximum number of deep supervision steps every time to avoid these issues, the Q-head becomes useless (assuming you don't want to use it during inference).
I really appreciate any help anyone can give me in understanding how they get around this, either in the original paper or in practice in more recent research.