Video

5 topics of this class 1) Supervised Learning 2) Machine Learning Strategy 3) Deep Learning 4) Unsupervised Learning 5) Reinforcement Learning

The audio is so bad..

Sincere students and smart

19:47 so is there a functional difference between calling the usage of softmax `attention` instead of the simpler word `search` beyond trying to be catchy?

It’s mind blowing how much overlap there is between startup info ( lean start up & y combinator ) & design thinking.

0:00:30 Review: Excess risk bound; Uniform convergence for 1) finite hypothesis class √{log|H|/n} 2) with p parameters √{p/n}; complexity measures of hypothesis 0:03:50 Limitations of these simple bounds 0:06:40 This lecture: More elaborate bounds: based on √{complexity(Θ,P)/n} 0:10:55 Def: Rademacher complexity, RC or R_n(H), a bound for uniform convergence 0:13:50 Correlation of f and σ: output {f(z_i)} and {σ_i}, loss fn and σ 0:15:00 High complexity function family F 0:18:55 Theorem E_{~data}[sup_{h \in H}(training loss - population loss) ≤ R_n(H)] 0:24:35 Interpretation of Rademacher complexity: How well the losses can correlate w random data 0:25:30 Binary classification example, 0-1 loss: l_01 0:32:15 Still binary, R_n(F)=1/2 R_n(H) How well can H memorize randomized labels: h(x_i)=σ_i 0:35:25 Intuition on R_n(F)=1/2 R_n(H) (true only for binary w 01 loss) 0:38:50 Connection btw RC and degrees of freedom 0:39:45 How do you think about the family of hypotheses and losses. Same cardinality, but very different output, non-linear loss fns (eg exponential) 0:42:20 Symmetrization proof 0:56:35 Removed expectation, removed randomness beyond that of the σ_i randomness 1:01:25 RC can depend on P distribution 1:09:35 Empirical Rademacher complexity ERC R_S(F) for dataset S, concentration around R_n(F) with high probability 1-δ 1:13:35 Theorem R_n(F) bounded by ERC + √{log δ/n} with high probability whp 1:15:45 Remark 1: logδ term is tiny compared to ERC 1:17:15 Remark 2: Rademacher complexities are translation invariant

The most funniest part comes when prof talks about booking the flight 😂😂

Hello Everyone! Thank you very much for uploading these materials. Cheers

Pity the resolution is so low, it is hard to read the code.Great content nevertheless!

AMazing stuff! Thank you for publishing this valuable material!

0:00:15 Rademacher Complexity RC ≤ formula for linear model; classification; RC bound 0:01:35 Binary classification setup 0:03:15 Finite hypothesis: Theorem: ∀f ∈ F bounded by M, then RC R_S(F)≤√{2M^2log(|F|)/n} on S={z_1,...,z_n} is bounded 0:07:15 l_{01}=l_γ((x,y),h)=𝟙(y sgn[h(x)]≤0) loss; signum issue → margin idea to resolve sign issue; apply to linear model 0:09:15 Margin intuition: Zero training error, perfect separation, can define margin(x) = y h_θ(x); dataset margin γ=min_i{y^(i) h_θ(x^(i))} 0:16:15 Margin dependent loss: l_γ=l_γ((x,y),h)=l_γ(yh(x)), ramp/margin loss, γ margin region/size 0:21:45 Ramp loss fn ≥ 0-1 loss fn: l_γ ≥ l_{01} 0:24:10 Test loss L_γ ≤ training loss \hat L_γ + 2 R_S(F). Look for bounds on R_S(F)! F=loss fn space indexed by h_θ ∈ H 0:26:20 Talagrand's lemma R_S(F) vs R_S(H): R_S(l_γ•H)≤1/γ R_S(H), where l_γ is γ-Lipschitz 0:48:30 Test loss bound is invariant to scaling, RC/γ_min 0:51:50 Linear model's RC: R_S(H)=BC/√{n}, with H={h(x) = w^T x} and ||x_i||_2<C, ||w||_2<B 1:07:25 Linear model's RC: R_S(H)=BC√{lg(d)/n}, with H={h(x) = w^T x} and ||x_i||_\infty<C, ||w||_1<B

very informational

Here's a engaging summary of the CZcams lecture you linked, focusing on the key highlights of Natural Language Understanding (NLU): **Natural Language Understanding: A Look Back and a Leap Forward 🚀** * **The Exciting Evolution of NLU:** Professor Manning kicks off the lecture by highlighting how far Natural Language Understanding has come. Starting from 2012's early wave of interest to today, the field has exploded! This makes it a super interesting time to study NLU, as things are changing fast! * **NLU's Amazing New Skills:** The things we can do with language models today are honestly mind-blowing. They can: * Write different styles of text (think poems, code, scripts... you name it!) ✍ * Translate between languages like a pro 🌎 * Answer your questions in actually helpful ways 🤔 * **The Tricky Question Test:** One way to see NLU progress is to ask the same question to models over time. The question "Which U.S. states border no U.S. states?" is surprisingly hard due to the word 'no'. Here's how models tackled it: * 1980: SHRDLU, a classic system, could reason about complex situations...but failed on basic things outside its knowledge 🙅‍♂ * 2009: Wolf... pretty much just listed all U.S. states 😅 * 2020-2021: OpenAI models start getting the idea of 'no other states', but make some geographical blunders (Puerto Rico isn't a state!) * 2022: DaVinci nails it! Alaska and Hawaii are the champs of isolation 🥇 * **Challenges and the Future:** This is just a taste of NLU's journey. There are still things these models struggle with, but the field is moving rapidly. It's an awesome time to be involved in shaping where NLU goes next! ✨

41:25 Lemma 1 Maximizing margin w/ l_2 regularization 51:55 Lemma 2 Maximizing margin w/ regularizing cross-entropy loss 58:55 Connection to l_1-SVM 1:12:25 Tools for multi-layer networks, intro

3:13 Into to this lecture: Test loss ≤ poly(Lipschitzness of f(x) on x1,...,xn; norm of θ) 4:14 Classical uniform convergence: 2 types 6:16 Union bound 7:15 Data dependent generalization bound 8:18 Data as regularizer 17:18 Generalized margin 21:53 Infinite covering number 23:11 Loss on f is bounded by complexity/(min gen margin * sqrt n) 38:07 Intuition: linear model, normalized margin, perturbation to cross the separation plane 41:03 All-layer margin, perturbation in each layer 46:55 Theorem L_{01}(f) loss is bounded O(sum of ||w_i||_{1,1}/(min gen margin * sqrt n)) 52:03 Lemma (decomposition) 56:32 Corollary 58:18 Proof 1:19:18 Compare w/ Bartlett etal '17 1:22:48 SGD prefers Lipschitzness on data points, implicitly maximizing all-layer margin 1:25:00 SAM Sharpness Aware Regularization: Lipschitzness on param, not on latent 1:27:20 Average margin test error, sum of complexities of each layer

4:35 Non-convex optimization, local minima, global minima 8:20 Neural Tangent Kernel NTK approach - Taylor-expansion (to linear) of f_θ(x) around θ^0 ≈ g(x) 14:25 Tangent feature map, tangent kernel 17:48 Extent of the validity of the linearization 20:57 When linearization works: 1,2,3 - Idea of good neighborhoods B(θ_0) 30:40 The n tangent features, one at each x data points. Mean squared loss of linearized f(x) 33:32 Lemma: What neighborhoods are good? Based on lowest singular value of tangent matrix. 39:00 Lemma: β-Lipschitzness in θ of ∇_θ f_θ(x) affects size of neighborhood 48:00 β/σ^2 is not scale invariant in the scale in f(x) 49:05 Case 1: Re-parametrize model with a scalar multiplier 58:22 Case 2: Overparametrize model 1:15:20 Key quality: more β-Lipschitz, ie, smoother in β, as more m neurons 1:21:50 Remark on online gradient descent

59:45 Theorem: NTK kernel sample efficiency vs NN sample efficiency

dead zone 💀

When i am listening to these lectures alone in my room and my office, and when the professor says "good good good question with a unique energy", it makes me with a silent smile. This man is amazing!!!!!.

Re: 2 vs 3 in the distillation objectives slide, he doesn't make it super clear, but 2 (output scores) refers to the softmax-produced probability vectors, and 3 refers to the vector of raw logits pre-softmax.

He's talking about Dungeon Keeper, which was a great game, I look down and see the likes at "666" - hilarious 😂 Great lecture!

Where are the professors?

Francis Galton's quote at the end and students clapping. Wow a wonderful lecture again! I am speeding through this like watching Netflix. Chris Piech is an awesome professor!

When the power goes out, where can i find a Stanford post-grad with masters degree to peddle a generator for me?

I love internet because I can learn from top universities

This is very very helpful, I would like to know if you have a course or if you teach product managements, as almost always PM courses share the same knowledge but in this 5 min video I was able to gain knowledge that I literally didn't gain in 2 hours course. THANK YOU <3

it would be great if CS25: V4 created another playlist in youtube.

Thank you sooooooo much and missssssss the psets eaaaaaaagerly.

Finally an explanation of the the acronym INTER PLANETARY FILE SYSTEM. amazing presentation!!!! Love the comment "if you do anything that will cause people to change their application system it will never be deployed"

Thank you again to stanford for making materials accessible online. Still working to fully grasp how the data is stored (I understand semantic triples) in terms of the actual programming languages used.

Be careful using anthropomorphic language when talking about LLMs. Eg: thoughts, ideas, reasoning. Transformers don’t “reason” or have “thoughts” or even “knowledge”. They extract existing patterns in the training data and use stochastic distributions to generate outputs.

Chris is an excellent lecturer! I love his sense of humor, it's a little absurd.

It is still lost on me as to why we would pass the training images to a RNN that outputs the learned parameters of an entirely different deep neural network (a MLP) that is useful in a future, similar task. The more straight forward method seems to be passing the training images directly to the MLP itself and having the parameters be learned that way. What am I missing here?

Gemini: This lecture is about prompting instruction fine-tuning and RLHF, which are all techniques used to train large language models (LLMs). LLMs are trained on a massive amount of text data and are able to communicate and generate human-like text in response to a wide range of prompts and questions. The lecture starts with going over zero-shot and few-shot learning, which are techniques for getting LLMs to perform tasks they weren't explicitly trained for. In zero-shot learning, the LLM is given a natural language description of the task and asked to complete it. In few-shot learning, the LLM is given a few examples of the task before being asked to complete a new one. Then the lecture dives into instruction fine-tuning, which is a technique for improving the performance of LLMs on a specific task by fine-tuning them on a dataset of human-written instructions and corresponding outputs. For example, you could fine-tune an LLM on a dataset of movie summaries and their corresponding reviews to improve its ability to summarize movies. Finally, the lecture discusses reinforcement learning from human feedback (RLHF), which is a technique for training LLMs using human feedback. In RLHF, the LLM is given a task and then asked to complete it. A human expert then evaluates the LLM's output and provides feedback. This feedback is then used to improve the LLM's performance on the task. The lecture concludes by discussing some of the challenges and limitations of RLHF, as well as the potential future directions for this field. One challenge is that it can be difficult to get humans to provide high-quality feedback, especially for complex tasks. Another challenge is that RLHF can be computationally expensive. However, RLHF is a promising technique for training LLMs to perform a wide range of tasks, and it is an area of active research.

😂 wt a joke

9:20 - BERT 33:17 - RoBERTa 34:40 - XLNet 38:28 - ALBERT 41:18 - T5 43:00 - ELECTRA

future artificial intelligence i was into talk this probability challenge Gemini ai talking ability rapid talk i suppose so it's splendid

just do it

❤

❤

how are there no comments on this

Very cool! Thanks for posting this publicly, it's really awesome to be able to audit the course :)

it's finally released! hope y'all enjoy(ed) the lecture 😁

Thanks for sharing this course and palestry Staford. Congratulations . Here the Brazil

Stanford's struggles with microphones continue.

that's 12 year ago, if you held $100 nvidia stock, now it will be $200,000+

☺️☺️☺️🥰🥰🥰

Awesome, thank you Stanford online for sharing these amazing video series

What are the pre requirements of this course?