Efficient Computing for Deep Learning, Robotics, and AI (Vivienne Sze) | MIT Deep Learning Series

Lecture by Vivienne Sze in January 2020, part of the MIT Deep Learning Lecture Series.

LECTURE LINKS:
Book coming out in Spring 2020!

OUTLINE:
0:00 - Introduction
0:43 - Talk overview
1:18 - Compute for deep learning
5:48 - Power consumption for deep learning, robotics, and AI
9:23 - Deep learning in the context of resource use
12:29 - Deep learning basics
20:28 - Hardware acceleration for deep learning
57:54 - Looking beyond the DNN accelerator for acceleration
1:03:45 - Beyond deep neural networks

CONNECT:
- If you enjoyed this video, please subscribe to this channel.

[lexfridman]

I really enjoyed this talk by Vivienne. Here's the outline:
0:00 - Introduction
0:43 - Talk overview
1:18 - Compute for deep learning
5:48 - Power consumption for deep learning, robotics, and AI
9:23 - Deep learning in the context of resource use
12:29 - Deep learning basics
20:28 - Hardware acceleration for deep learning
57:54 - Looking beyond the DNN accelerator for acceleration
1:03:45 - Beyond deep neural networks

[NomenNescio]

Thank you for sharing the lecture, this is the type of content I really enjoy.

[gonzalochristobal]

thank you lex, the amount of information you already shared is invaluable, eternally grateful

[JonMcGill]

I used to be a Field Apps engineer for telecom, and she's certainly correct about the power problem with respect to chip technology. Very likeable lecturer!!

[colouredlaundry]

I am not an expert in the field of Vivienne Sze, however, she was an extremely good lecturer. Every concept was extremely clear.

[jayhu]

Many thanks for sharing to the people that not study can afford at the MIT. Respect.

[samuelec]

Impressive amounts of information delivered by this lady!. To watch a such high densely packed informative video I had to take more than few breaks.
I wonder how she managed to go through those 80 slides so fast and if there is someone that watched it all in one go without lose the focus !

[pierreerbacher]

The density of neurons in this channel is incredibly high.

[nikhilpandey]

I was researching about this on my own. I have been doing the network pruning wrong. I wouldn’t mind a hit in accuracy if my latency budget were met but now I think I can be far more frugal with the decrease in accuracy. Thanks a lot.

[JousefM]

Thanks for a rather "exotic" topic I need to learn about as an AI newbie, much appreciated Lex!

[summersnow]

Excellent lecture 👏👏👏. Things that we don’t usually think about as a ML practitioner but highly important. Great insights.

[warsin]

I love this I will rewatch everything when I'm older and hopefully understand better and deeper I'm only a junior in high school 😖

[XCSme]

Great video and an interesting problem.
Why stop at architecture? What about using different materials for specialized DNN hardware? Maybe using some lower power transistors that are less accurate but good enough for inference. I don't think the brain neurons are always 100% accurate and consistent, but the brain seems to be somewhat fault tolerant.

[BlackHermit]

FastDepth is really interesting. Could be useful for many people.

[machinimaaquinix]

This was a great talk, thank goodnees Youtube has a .75 speed mode. She talks fast!

[valfredodematteis-poet]

Thank you, every video you post is incredibly useful. Though, it is really hard to enter this field from scratch: everything you learn forces you to go learn thousands other things, it gets really frustrating sometimes. I hope in time this will go better

[masbro]

1:10:37 its 100x faster than FPGA ?? wth, wow, thats blowing my mind, i thought designing custom hardware for specialized algorithm on FPGA its the fastest way on the planet, is it really??

[adamsimms]

I'm trying to imagine how this structure will become half relevant as we move into UltraRAM which is as close to as fast as DRAM but NOT volatile like memory stick type RAM. What are the implications if the data can be laid out and accessed in place where it is saved. Suddenly the whole structure is no longer useful.

[punyaslokdutta]

Trade off between number of filters on the 3D Convolution and a 4D Convolution ? . Convolution is a matrix operation (w*Imap+ Bias). RELU Activation is mostly used to provide non-linearity . I feel the number of Filters is needed to see higher abstracted stuff . For instance, The initial layer of a CNN understands pixels based information primarily for edges, cuts, depths. The layer following it understands shape, structures. The further layers help us understand semantic meaning of eyes, skin, ears, nose, face. But, How will the model perform when instead of multiple filters, we having more layers . That is, the information in filters in stuffed inside the CNN layers. Or is it done for easing computation while training ?

[minhongz]

So essentially power consumption and speed are almost equivalent for AI chips. Does anyone know what architecture Tesla chips employ?