วีดีโอ

When mentioning KV-cache, Lucas incorrectly highlights the sequence length as the embedded vector length, when the latter is the d_model, and sequence length is the context window

at 37:00, why are the special tokens different for the two sequences?

❤❤

Fourier transform is mentioned. How does one learn the applications of that in the context of MLL?

please make it 2k or 4k

beyer 😍😍😍

tnx for sharing

As an AI expert and teacher, I love the way you present the threats and challenges of AI in a way that has room for hope. I ordered your book and will likely use it in my Machine learning course to let computer science engineering students reflect on the future of AI. If ever you're in Belgium, I'd love to invite you for a guest lecture at Ghent University, from which many students (and professors) would surely benefit!

RL should thank Schopenhauer for this idea. Intellect is just servant to the Will i.e. will to live and procreate.

Artificial Intelligence is a Discovery not an invention. I was showing someone the AI bots that learned to play soccer. They stated "But when I play soccer , I know i am playing soccer." and so my response was. "Well you think that, because your Brain gives you a reward when you do." Meaning is a Human illusion.

I believe the figures around 13:00 and on the next slide are misleading. To my understanding the attention weights softmax(QK^T) would always be a square matrix TxT matrix with T the sequence length. Also the dimensionality of the keys and queries needs to be the same, otherwise the inner product in softmax(QK^T) would have a shape mismatch.

Thank you i loved it. And now there is a new book from him available on the topic: “Deep Learning”.

I didn’t expect fish to have more synapses then rats.

Batchnorm can be statefull, but most optimizers should be able to fuse the affine transformation into the convolution layer before it. In that case BN has no computation cost at runtime

nice presentation for beginners.

exciting to find this on yt!

How can I add 'Self-Attention mechanism' to Vector Quantized Diffusion Model for Text-to-Image Synthesis?

Really good lecture, thanks for uploading!

Thanks

Great course. Great human being.

Very very good overview, thanks for sharing!

This was the best overview of transformers I’ve seen to date. Great talk!

th-cam.com/video/-hcxTS5AXW0/w-d-xo.html q_{\phi} instead of q_{\theta}?

Good presentation 🎉

Poor audio quality.

þrðmð§m 🌸

Essentially, reward is enough is just saying optimizing an objective is enough. That is definitely plausible, provided there is sufficient expressivity of the system to learn complex behaviour needed for the objective in the environment. The only issue I have is that the reward signal may be sparse, or stochastic, which would require many runs of the environment to learn it. How then, are humans/animals able to learn from limited trials and not kill ourselves in the process. Imagine if humans learn how to drive cars like how a DQN / A3C agent learns. There would be a million car crashes before the car could get out of the car park. Hence, reward is good and important - who drives the reward (external or internal) is a good point to think about. Could some behaviours also be innate?

Excellent webinar

Has cognitive 'science' inhibited the development of AI?

Thank you for sharing, insightful!

Vim após assistir o podcast com Serjão é muito bom saber que você está avançando no projeto boa sorte!

it is possible to update full course ty!

This kind of coverage of a published paper, especially this paper, is extremely helpful. Why this has only 3.6k views I cannot say. This one is big.

Thanks for sharing.

Thanks for posting this professor Cartis, I look forward to watching these lectures.

12:19 If a robot maximising only cleanliness is around children, the children are in trouble

I find this hypotheses very plausible, but there needs to be a better effort at defining the working terms, such as intelligence. Otherwise, there is really no argument.

9:40 I didn't know pain is a reward :D

He literally told "reward is enough, but actually not enough".

thank you Andrea and organizers. Insights exposed are top level, please consider that the loop mentioned also scale up as a meta-analogy and is really amazing how one promising tool for understand the whole complexity of the brain is clearly an emulation of its proper function.. to investigate itself, ourselves. A friend of mine is a "language key" believer, and this presentation points that.

Is it possible to provide a link for the slides of this talk? Thanks a lot.

Sorry ' Do you use Matlab or python as your framework to apply your methodology ?