Thank you so much MIT and instructors for making these very high quality lectures available to everyone. Students from developing countries who have aspirations to achieve something big is now possible with this type of content and information!
Over all videos on RU-vid that explained about Transformer architecture (including the visual explanation) , this is the BEST EXPLANATION ever done. Simple, contextual, high level, step by step complexity progression. Thank you the educators and MIT!
@@ukaszkasprzak5921 Kognitywistyka UW Zagadnienia z AI, machine learningu i matematyki są tu omawiane obok zagadnień humanistycznych: Lingwistyka, Filozofia Umysłu, Psychologia Poznawcza etc. Radzę przejrzeć Program studiów, proste googlowanie wystarczy
As a CS student from University of Tehran, you guys don't have any idea how much such content could be helpful and the idea that all of this is free make it really amazing. Really appreciate it Alexander and Ava. Best hops.
I watched and read a lot of content about Transformers and never understood what are those three Q, K, and V vectors doing so I coulnd't understand how attention works, until today when I watched this lecture doing the analogy of RU-vid search and the Iron Man picture. Now it became much much clearer! Thanks for the brilliant analogies that you are making!
This is my favorite subject :) (following is self clarification of said words that feel exaggerated) 4:08 - binary classification or filtering is a sequence of steps: - new recording - retrieval of a constant record - compare new and constant record - express a property of the compare process So, sequencing really is a property of maybe all systems. While "wave sequencing" is built on top of a Sequencer System, that repeatedly uses the "same actions" per sequence element.
Wow, Transformers, and Attention was an absolute lifesaver! 🚀🙌 The explanations were crystal clear, and I finally have a solid grasp on these concepts. This video saved me so much time and confusion. Huge thanks to the Ava for making such an informative and engaging tutorial! Can't wait to delve deeper into the world of AI and machine learning. 🤖💡
Indeed commendable the way this lecture has been ordered and difficult topic like self-attention has been lucidly explained. Thanks to the instructors, really appreciated.
This is what we need in this day and age, the teaching is amazing and can be understood by people of variable intelligence. Nice work and thanks for this course.
I always meant to watch these lectures since 2020, but something always comes up. Now, nothing is going to stop me. Not even nothing. Great lectures, best way to learn.
Same man. The academic stress as an undergraduate was my "something always comes up," but since I just graduated a few days ago, I now have no excuse to not indulge myself in these videos lol.
Came here to refresh my memory of deep learning for sequential data. I really like how Ava brings us from one algorithm to another. It makes perfect sense to me.
Summary by Gemini: The lecture is about recurrent neural networks, transformers, and attention. The speaker, Ava, starts the lecture by introducing the concept of sequential data and how it is different from the data that we typically work with in neural networks. She then goes on to discuss the different types of sequential modeling problems, such as text generation, machine translation, and image captioning. Next, Ava introduces the concept of recurrent neural networks (RNNs) and how they can be used to process sequential data. She explains that RNNs are able to learn from the past and use that information to make predictions about the future. However, she also points out that RNNs can suffer from vanishing and exploding gradients, which can make them difficult to train. To address these limitations, Ava introduces the concept of transformers. Transformers are a type of neural network that does not rely on recurrence. Instead, they use attention to focus on the most important parts of the input data. Ava explains that transformers have been shown to be very effective for a variety of sequential modeling tasks, including machine translation and text generation. In the last part of the lecture, Ava discusses the applications of transformers in various fields, such as biology, medicine, and computer vision. She concludes the lecture by summarizing the key points and encouraging the audience to ask questions.
I come back every year to check these lectures and to see what innovations made it into the lectures. Pleasantly surprised to see the name change, congrats!
Thank you very much for this great oppurtunity to watch MIT lectures. always dreamt of a world class education and finally im doing a degree in AI and such videos are supporting my learning process very much
I already have some knowledge on the subject, however, I like to keep myself updated and there is always something new to learn. She clearly explains how what she is teaching really works. The whole video is worth watching.
If someone is looking for an easy way to understand transformer architecture, this lecture is for you. Amazing job. Thanks for sharing it as open source :p
Great material and the best educator!. Thank you for the fantastic video! The material was not only informative but also engaging, and the quality of the presentation was top-notch. Your depth of knowledge truly shines through, making the learning experience both enriching and enjoyable. Presented such complex material with such ease. You've done an exceptional job in communicating the concepts clearly. Great work!" and everything is free! Great job MIT team!!
I just started learning about RNN and LSTM especially for NLP and found this video very helpful to me. It would be really exciting if you provided a video about transformers in more depth :)
🎯Course outline for quick navigation: [00:09-02:02]Sequence modeling with neural networks -[00:09-00:37]Ava introduces second lecture on sequence modeling in neural networks. -[00:55-01:46]The lecture aims to demystify sequential modeling by starting from foundational concepts and developing intuition through step-by-step explanations. [02:02-13:24]Sequential data processing and modeling -[02:02-02:46]Sequential data is all around us, from sound waves to text and language. -[03:10-03:50]Sequential modeling can be applied to classification and regression problems, with feed-forward models operating in a fixed, static setting. -[05:02-05:26]Lecture covers building neural networks for recurrent and transformer architectures. -[11:56-12:37]Rnn captures cyclic temporal dependency in maintaining and updating state at each time step. [13:24-20:04]Understanding rnn computation -[14:40-15:04]Explains rnn's prediction for next word, updating state, and processing sequential information. -[15:05-15:47]Rnn computes hidden state update and output prediction. -[16:17-17:05]Rnn updates hidden state and generates output in single operation. -[18:45-19:39]The total loss for a particular input to the rnn is computed by summing individual loss terms. the rnn implementation in tensorflow involves defining an rnn as a layer operation and class, initializing weight matrices and hidden state, and passing forward through the rnn network to process a given input x. [20:05-29:13]Rnn in tensorflow -[20:05-20:54]Tensorflow abstracts rnn network definition for efficiency. practice rnn implementation in today's lab. -[21:16-21:43]Today's software lab focuses on many-to-many processing and sequential modeling. -[22:53-23:21]Sequence implies order, impacting predictions. parameter sharing is crucial for effective information processing. -[25:04-25:29]Language must be numerically represented for processing, requiring translation into a vector. -[28:29-28:56]Predict next word with short, long, and even longer sequences while tracking dependencies across different lengths. [29:14-41:53]Rnn training and issues -[30:02-30:27]Training neural network models using backpropagation algorithm for sequential information. -[30:45-31:43]Rnns use backpropagation through time to adjust network weights and minimize overall loss through individual time steps. -[32:03-32:57]Repeated multiplications of big weight matrices can lead to exploding gradients, making it infeasible to train the network stably. -[35:45-37:18]Three ways to mitigate vanishing gradient problem: change activation functions, initialize parameters, use a more robust version of recurrent neural unit. -[36:13-37:01]Relu activation function helps mitigate vanishing gradient problem by maintaining derivatives greater than one, and weight initialization with identity matrices prevents rapid shrinkage of weight updates. -[37:54-38:25]Lstms are effective at tracking long-term dependencies by controlling information flow through gates. -[40:18-41:13]Build rnn to predict musical notes and generate new sequences, e.g. completing schubert's unfinished symphony. [41:53-50:11]Challenges in rnn and self-attention -[43:58-44:40]Rnns face challenges in slow processing and limited capacity for long memory data. -[46:37-47:00]Concatenate all time steps into one vector input for the model -[47:21-47:45]Feed-forward network lacks scalability, loses in-order information, and hinders long-term memory. -[48:11-48:34]Self-attention is a powerful concept in deep learning and ai, foundational in transformer architecture. -[48:58-49:25]Exploring the power of self-attention in neural networks, focusing on attending to important parts of an input example. [50:13-56:20]Neural network attention mechanism -[50:13-50:43]Understanding the concept of search and its role in extracting important information from a larger data set. -[51:52-55:24]Neural networks use self-attention to extract relevant information, like in the example of identifying a relevant video on deep learning, by computing similarity scores between queries and keys. -[53:32-53:54]A neural network encodes positional information to process time steps all at once in singular data. -[55:32-55:57]Comparing vectors using dot product to measure similarity. [56:20-01:02:47]Self-attention mechanism in nlp -[56:20-57:14]Computing attention scores to define relationships in sequential data. -[59:11-59:39]Self-attention heads extract high attention features, forming larger network architectures. -[01:00:32-01:00:56]Self-attention is a key operation in powerful neural networks like gpt-3. offered by Coursnap
00:16 Building neural networks for handling sequential data 03:19 Sequential data introduces new problem definitions for neural networks 10:03 Recurrent Neural Networks link computation and information via recurrent relation. 13:37 RNN processes temporal information and generates predictions. 20:22 Key criteria for designing effective RNNs 23:33 Recurrent neural networks design criteria and need for more powerful architectures. 30:08 Back propagation through time in RNN involves back propagating loss through individual time steps and handling sequential information. 33:23 Vanishing gradient problem in recurrent neural networks 40:03 RNNs used for music generation and sentiment classification 43:32 RNNs have encoding bottlenecks and processing limitations 49:45 Self-attention involves identifying important parts and extracting relevant information. 52:51 Transformers eliminate recurrence and capture positional order information through positional encoding and attention mechanism. 59:35 Self-attention heads extract salient features from data. 1:02:49 Starting work on the labs
This is the best lecture on RU-vid! Thank you for the clear explanation. I wish you could delve deeper into the transformer architecture, though, as it was only covered in the last 15 minutes. Nevertheless, this is the most understandable video on the topic. I've watched nearly all of them, but this one stands out as the best! It would be great if you provided a more detailed explanation of transformers.
Awsome! Video!! Very well thought out lecture. Keep rockin' !!! You just solved my problem in my NNW optimization project, in just two sentences.🤣 For 4 months, this has been driving me completely insane.💥🤣🔫 I think I'm in love.😀
Ava I don't think you understood the problem of gradient explosion, you explained it really bad, an evident drop of quality passing from the alexander lesson to this
