Yichao Cai

PhD Student @ Australian Institute for Machine Learning

[Yichao Cai @YichaoCai]

@jchencxh Theoretically, we may need the right equivalences and distinctions; practically, yeah, whether those actually get learned also depends on the learning interface/architecture.

[Yichao Cai @YichaoCai]

@jchencxh Thanks, this helps a lot. Maybe on the supervision side, hierarchy filling needs both: a stable target that anchors varying lower-level realizations, and enough contextual variation to make levels distinguishable.

[Yichao Cai @YichaoCai]

@jchencxh More generally, do you think that ``filling the hierarchy'' requires varying the relationships among semantic factors across prediction contexts so that different levels become separately identifiable?

[Yichao Cai @YichaoCai]

@jchencxh Good blog. I really like the hierarchy-filling perspective. Could another way to think about it be that learning abstractions at different levels requires supervision to induce enough variation so that those levels have distinguishable predictive consequences?

[Yichao Cai @YichaoCai]

@jchencxh E.g., could MAE’s limitation be that random masking in pixel space often doesn’t sufficiently separate high-level semantic explanations from lower-level ones, since ``p(masked pixels∣visible pixels)'' can often be predicted well without resolving the higher-level structure?

[Yichao Cai @YichaoCai]

AI sycophancy becomes critically dangerous when it corrodes the epistemic integrity of humanity.

[Yichao Cai @YichaoCai]

@overleaf Relax, nothing is in control. Let's have a rest.😅

[Yichao Cai @YichaoCai]

https://t.co/F7UHzq0xkE

[Yichao Cai @YichaoCai]

Don't drink coffee at 10 PM. 🙃

[Yichao Cai @YichaoCai]

Introducing my "LlaMA model". 👀

[Yichao Cai @YichaoCai]

学如逆水行舟，不进则退；心似平原跑马，易放难收。😮‍💨

[Frank (Haofan) Wang @Haofan_Wang]

Image-based stylization has been supported in InstantStyle. Found more information at https://t.co/7lhohKTTms…. We will further combine InstantID for face stylization once stars reach 1K on GitHub.

[🐢🐢🐢John Green🐢🐢🐢 @johngreen]

Today is #WorldTBDay. Around a billion people have died of tuberculosis since 1800. At the current rate of decline, we can expect another 100 million to die of TB in the next century. Each of those deaths would be needless.
It all reminds me of a different date: October 24th.

[Yichao Cai @YichaoCai]

@theemadwoman Red -> All too well

[Yichao Cai @YichaoCai]

The server crashed again. 🤡

[Yichao Cai @YichaoCai]

🥲🥲🥲

[Overleaf @overleaf]

⚠️ Attention: ⚠️ We have mitigated a denial of service attack. The service should be back for logged in users. The homepage and login page are still down. Our team is working to bring the service back up as fast as possible: https://t.co/y7aRh5SBN4 Sorry for the inconvenience

[Yichao Cai @YichaoCai]

@preskill @overleaf Intermittent Server Error.😮‍💨

[Yichao Cai @YichaoCai]

@overleaf 👏

[Yichao Cai @YichaoCai]

@baabysharkdoo @overleaf God bless you.

[Yichao Cai @YichaoCai]

Even though I know these servers have many backups, I'm still worried about losing my data (those LaTeX formulas and tables take too much time to organize).🤡