- 13
- 17 307
Anastasia Borovykh
United Kingdom
Приєднався 26 сер 2020
I make videos on machine learning. Watch, like, subscribe, give feedback, share ideas! :)
Demystifying Large Language Models in 45 minutes (non-technical)
What was planned to be a 30 minute non-technical introduction into the core components of Large Language Models turned into a 45 minute one... I dive into how to build these models: the model architecture, the training datasets and all the post-training that goes into them. I then discuss some of their awesome abilities, including their use in agents, but also the challenges (think hallucinations, sycophancy) that they still face.
Share, like, subscribe (jk) or just let me know if you have any questions ;)
Share, like, subscribe (jk) or just let me know if you have any questions ;)
Переглядів: 503
Відео
Steering vectors: tailor LLMs without training. Part II: Code (Interpretability Series)
Переглядів 6254 місяці тому
See Part I for an intro into Steering Vectors ua-cam.com/video/cp-YSyc5aW8/v-deo.html. Code from this video: github.com/abrvkh/explainability_toolkit/blob/main/notebooks/phi3_steering_vectors.ipynb State-of-the-art foundation models are often seen as black boxes: we send a prompt in and we get out our - often useful - answer. But what happens inside the system as the prompt gets processed remai...
Steering vectors: tailor LLMs without training. Part I: Theory (Interpretability Series)
Переглядів 1,3 тис.4 місяці тому
State-of-the-art foundation models are often seen as black boxes: we send a prompt in and we get out our - often useful - answer. But what happens inside the system as the prompt gets processed remains a bit of a mystery & our ability to control or steer the processing into specific directions is limited. Enter steering vectors! By computing a vector that represents a particular feature or conc...
Decoding hidden states of Phi-3 with LogitLens (Interpretability Series)
Переглядів 5556 місяців тому
In this series we discuss interpretability and explainability methods for deep neural nets focussing on Microsoft's Phi3. In this episode we zoom into LogitLens, a simple method that allows to decode intermediate representations. Blog post: www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit-lens Language paper: arxiv.org/pdf/2402.10588 Notebook: github.com/abrvkh/explainabilit...
State Space Models (S4, S5, S6/Mamba) Explained
Переглядів 6 тис.8 місяців тому
In this video we give a quick(ish) overview of state space models and how to use them as a layer in a neural network. We cover S4, S5 and S6/Mamba. References I like: - S4: arxiv.org/abs/2111.00396, stacks.stanford.edu/file/druid:mb976vf9362/gu_dissertation-augmented.pdf, srush.github.io/annotated-s4/ - S5: arxiv.org/abs/2208.04933 - S6/Mamba: arxiv.org/abs/2312.00752 - Mamba as attention: arxi...
Influence functions for large language models - why LLMs generate what they generate
Переглядів 1,5 тис.Рік тому
Influence functions can be used to measure how the model parameters & output is affected by the addition or removal of a datapoint to the train dataset. In this video we discuss how influence functions can be used to understand why large language models generate what they generate. If you have any questions, let me know in the comments :) References: - Original paper: arxiv.org/pdf/2308.03296.p...
Three times artificial neural networks are nothing like the human brain (+ are they ever alike?)
Переглядів 601Рік тому
In this video we discuss three differences between how the human brain and artificial neural networks process information. We first show some similarities between activations in ANNs and measurements taken from the brain. We then discuss adversarial examples, generalization gaps and data reconstruction. Let me know if you see any typo's or if you have any questions :)
Does ChatGPT memorize train data? - exploring memorization in neural networks
Переглядів 374Рік тому
In this video I discuss some papers on memorization in neural networks & generative models and we then explore whether ChatGPT memorizes datapoints. Memorization in my definition means that the model "remembers" relatively rare datapoints. Of course, the discussion can be on whether this kind of behavior is actually wanted for ChatGPT; one could say anything that is on the internet should be se...
Bounding the generalisation error in machine learning with concentration inequalities
Переглядів 7032 роки тому
In this video we briefly describe how we can bound the generalisation error (defined as the true risk minus the empirical risk) using concentration inequalities. If you see any typo's or have any questions - let me know :)
A very, very basic coding tutorial for distributed optimization
Переглядів 1,1 тис.2 роки тому
In this video we will code up a basic distributed optimizer for minimizing a sum of quadratic objectives. Let me know if there are any comments or questions - THANK YOU FOR WATCHING :)
A very, very basic introduction into distributed optimization
Переглядів 3,2 тис.2 роки тому
A very brief intro into distributed optimization where we discuss the setup of the problem, a distributed optimization algorithm and some intuition into the convergence. Typo's (comment if you find more!) - In "a quick convergence proof", V_t should have a factor 1/2 in front of it; in dV_t/dt there should be no dt on the right-hand side, and in the final dV_t/dt bound it should be 2\mu_fV_t.
Efficient distributed optimization with mirror descent + a mirror descent introduction
Переглядів 5073 роки тому
In this video we discuss the mirror descent algorithm in a distributed setup. Starting with a brief introduction into mirror descent and its relation to preconditioning we discuss how to define a distributed variant of the algorithm that converges both fast and close to the optimizer.
To interact or not? The convergence properties of interacting stochastic mirror descent.
Переглядів 4644 роки тому
In this video we discuss the benefits of running multiple copies of a gradient-based optimizer, which we refer to as particles, and letting these interact. This video first appeared at the ICML Workshop "Beyond first-order methods in ML systems". Anastasia Borovykh; joint work with Panos Parpas, Nikolas Kantas and Greg Pavliotis