6.

S094: Deep Learning for Self-Driving Cars

Lecture 1: Introduction to Deep Learning and Self-Driving Cars
cars.mit.edu

Course 6.S094: Lex Fridman: Website: January

References: [13, 14] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Administrative
Website: cars.mit.edu
Contact Email: [email protected]
Required:
Create an account on the website.
Follow the tutorial for each of the 2 projects.
Recommended:
Ask questions
Win competition!

Course 6.S094: Lex Fridman: Website: January

Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Target Audience
You may be:
New to programming
New to machine learning
New to robotics

What you will learn:

An overview of deep learning methods:
Deep Reinforcement Learning
Convolutional Neural Networks
Recurrent Neural Networks
How deep learning can help improve each component of
autonomous driving: perception, localization, mapping, control,
planning, driver state
Course 6.S094: Lex Fridman: Website: January
Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Target Audience
Not many equation slides like the following:

* Though it would be more efficient, since the above is LaTeX code automatically
generated character by character with Recurrent Neural Networks (RNNs)

[35] Andrej Karpathy. The Unreasonable Effectiveness of Recurrent Neural Networks." (2015).
Course 6.S094: Lex Fridman: Website: January
References: [35] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Guest Speakers

Course 6.S094: Lex Fridman: Website: January

Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Project: DeepTraffic

Course 6.S094: Lex Fridman: Website: January

Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Project: DeepTesla

Course 6.S094: Lex Fridman: Website: January

Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Schedule

Course 6.S094: Lex Fridman: Website: January

Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Defining (Artificial) Intelligence
March 25, 1996
Special Purpose:
Can it achieve a well-defined finite
set of goals?
General Purpose:
Can it achieve poorly-defined
unconstrained set of goals?

(1995) (2002) (2009)

Course 6.S094: Lex Fridman: Website: January

References: [59, 60, 61] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Formal tasks: Playing board games, card
games. Solving puzzles, mathematical
and logic problems.
Expert tasks: Medical diagnosis,
engineering, scheduling, computer
hardware design.
Mundane tasks: Everyday speech,
written language, perception, walking,
object manipulation.

Course 6.S094: Lex Fridman: Website: January

References: [49] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 How Hard is Driving?

Open Question:
Is driving closer to chess or to everyday conversation?

Course 6.S094: Lex Fridman: Website: January

Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Chess Pieces: Self-Driving Car Sensors
External
1. Radar
2. Visible-light camera
3. LIDAR
4. Infrared camera
5. Stereo vision
6. GPS/IMU
7. CAN
8. Audio

Internal
1. Visible-light camera
2. Infrared camera
3. Audio

Course 6.S094: Lex Fridman: Website: January

References: [71] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Chess Pieces: Self-Driving Car Tasks

Localization and Mapping:

Where am I?

Scene Understanding:
Where is everyone else?

Movement Planning:
How do I get from A to B?

Driver State:
Whats the driver up to?

Course 6.S094: Lex Fridman: Website: January

References: [71] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 DARPA Grand Challenge II (2006)

Result: Stanfords Stanley wins

Course 6.S094: Lex Fridman: Website: January

References: [73] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 DARPA Urban Challenge (2007)

Result: CMUs Boss (Tartan Racing) wins

Course 6.S094: Lex Fridman: Website: January

References: [74] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Industry Takes on the Challenge

Waymo / Google Self-Driving Car Tesla Autopilot

Uber nuTonomy

Course 6.S094: Lex Fridman: Website: January

References: [75] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 If Driving is a Conversation:
How Hard is it to Pass the Turing Test?
1. Natural language processing to enable
it to communicate successfully
2. Knowledge representation to store
information provided before or during
the interrogation
3. Automated reasoning to use the stored
information to answer questions and to
draw new conclusions

Turing Test:
Can a computer be mistaken for a
human more than 30% of the time?

Course 6.S094: Lex Fridman: Website: January

Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Neuron: Biological Inspiration for Computation

Neuron: computational building (Artificial) Neuron: computational

block for the brain building block for the neural network
Human brain: (Artificial) neural network:
~100-1,000 trillion synapses ~1-10 billion synapses

Human brains have ~10,000 computational power than computer brains.

Course 6.S094: Lex Fridman: Website: January

References: [18] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Perceptron: Forward Pass

Course 6.S094: Lex Fridman: Website: January

References: [78] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Perceptron Algorithm

Provide training set of (input, output) pairs and run:

1. Initialize perceptron with random weights
2. For the inputs of an example in the training set, compute the Perceptrons output
3. If the output of the Perceptron does not match the output that is known to be
correct for the example:
1. If the output should have been 0 but was 1, decrease the weights that had an input of 1.
2. If the output should have been 1 but was 0, increase the weights that had an input of 1.

4. Go to the next example in the training set and repeat steps 2-4 until the Perceptron
makes no more mistakes

Course 6.S094: Lex Fridman: Website: January

References: [22, 78] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Neural Networks are Amazing

Universality: For any arbitrary function f(x), there exists a neural

network that closely approximate it for any input x

Universality is an incredible property!* And it holds for just 1 hidden layer.

* Given that we have good algorithms for training these networks.

Course 6.S094: Lex Fridman: Website: January

References: [62] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Special Purpose Intelligence

Course 6.S094: Lex Fridman: Website: January

References: [65] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Neural Networks are Amazing:
General Purpose Intelligence
Policy Network:

80x80 image (difference image)

2 actions: up or down
200,000 Pong games

This is a step towards general purpose

artificial intelligence!
Andrej Karpathy. Deep Reinforcement
Learning: Pong from Pixels. 2016.

Course 6.S094: Lex Fridman: Website: January

References: [63] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Neural Networks are Amazing:
General Purpose Intelligence

Every (state, action) pair is rewarded

when the final result is a win.
Every (state, action) pair is punished
when the final result is a loss.

The fact that this works at all is amazing!

It could be called general intelligence but not yet human-level intelligence

Course 6.S094: Lex Fridman: Website: January

References: [63] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Current Drawbacks
Lacks Reasoning:
Humans only need simple instructions:
Youre in control of a paddle and you can move it up and down, and your
task is to bounce the ball past the other player controlled by AI.
Requires big data: inefficient at learning from data
Requires supervised data: costly to annotate real-world data
Need to manually select network structure
Needs hyperparameter tuning for training:
Learning rate
Loss function
Mini-batch size
Number of training iterations
Momentum: gradient update smoothing
Optimizer selection
Defining a good reward function is difficult

Course 6.S094: Lex Fridman: Website: January

References: [63, 64] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Current Drawbacks
Defining a good reward function is difficult Coast Runners: Discovers local pockets of
high reward ignoring the implied bigger picture goal of finishing the race.

In addition, specifying a reward function for self-driving cars raises ethical questions
Course 6.S094: Lex Fridman: Website: January
References: [63, 64] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Deep Learning Breakthroughs: What Changed?
Compute
CPUs, GPUs, ASICs
Organized large(-ish) datasets
Imagenet
Algorithms and research:
Backprop, CNN, LSTM
Software and Infrastructure
Git, ROS, PR2, AWS, Amazon Mechanical Turk, TensorFlow,
Financial backing of large companies
Google, Facebook, Amazon,

Course 6.S094: Lex Fridman: Website: January

Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Useful Deep Learning Terms
Basic terms:
Deep Learning = Neural Networks
Deep Learning is a subset of Machine Learning
Terms for neural networks:
MLP: Multilayer Perceptron
DNN: Deep neural networks
RNN: Recurrent neural networks
LSTM: Long Short-Term Memory
CNN or ConvNet: Convolutional neural networks
DBN: Deep Belief Networks
Neural network operations:
Convolution
Pooling
Activation function
Backpropagation

Asimov Institute. A mostly complete chart of neural networks.

Course 6.S094: Lex Fridman: Website: January

References: [4] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Neural Networks: Proceed with Caution

Course 6.S094: Lex Fridman: Website: January

References: [50] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Deep Learning is Representation Learning

[20] Goodfellow et al. "Deep learning." (2017).

Course 6.S094: Lex Fridman: Website: January
References: [20] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Representation Matters

Course 6.S094: Lex Fridman: Website: January

References: [20] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Deep Learning is Representation Learning

Course 6.S094: Lex Fridman: Website: January

References: [20] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Deep Learning: Scalable Machine Learning

Course 6.S094: Lex Fridman: Website: January

References: [20] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Applications: Object Classification in Images

Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. "Imagenet classification with deep
convolutional neural networks." Advances in neural information processing systems. 2012.
Course 6.S094: Lex Fridman: Website: January
References: [4] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Pause: Progress on ImageNet

Course 6.S094: Lex Fridman: Website: January

References: [55] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Computer Vision is Hard:
Illumination Variability

Course 6.S094: Lex Fridman: Website: January

References: [66] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Computer Vision is Hard:
Pose Variability and Occlusions

Parkhi et al. "The truth about cats and dogs. 2011.

Course 6.S094: Lex Fridman: Website: January
References: [69] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Computer Vision is Hard:
Intra-Class Variability

Parkhi et al. "Cats and dogs." 2012.

Course 6.S094: Lex Fridman: Website: January

References: [70] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Pause: Object Recognition / Classification

Course 6.S094: Lex Fridman: Website: January

Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Pause: Segmentation

Original Ground Truth FCN-8

Source: Long et al. Fully Convolutional Networks for Semantic Segmentation. CVPR 2015.
Course 6.S094: Lex Fridman: Website: January
Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Pause: Object Detection

Course 6.S094: Lex Fridman: Website: January

Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Applications: Object Detection and Localization in Images

He Kaiming, Xiangyu Zhang, Shaoqing Ren, and Jian Su.

"Deep residual learning for image recognition." (2015).

Course 6.S094: Lex Fridman: Website: January

References: [31, 32] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Applications: Colorization of Images

Zhang, Richard, Phillip Isola, and Alexei A. Efros. "Colorful Image Colorization." (2016).
Course 6.S094: Lex Fridman: Website: January
References: [25, 26] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Applications: Automatic Translation of Text in Images

Google Translate
Course 6.S094: Lex Fridman: Website: January
References: [30] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 (Pause) Flavors of Neural Networks

Recurrent Neural Networks

Vanilla
Neural Andrej Karpathy. The Unreasonable Effectiveness
Networks of Recurrent Neural Networks." (2015).

Course 6.S094: Lex Fridman: Website: January

References: [35] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Applications: Handwriting Generation from Text

Input:

Output:

Alex Graves. "Generating sequences with

recurrent neural networks." (2013).

Course 6.S094: Lex Fridman: Website: January

References: [] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Applications: Character-Level Text Generation

Naturalism and decision for the majority of Arab

countries' capitalide was grounded by the Irish
language by [[John Clair]], [[An Imperial Japanese
Revolt]], associated with Guangzham's sovereignty.
His generals were the powerful ruler of the Portugal in
the [[Protestant Immineners]], which could be said to
be directly in Cantonese Communication, which
followed a ceremony and set inspired prison, training.

Andrej Karpathy. The Unreasonable Effectiveness of

Recurrent Neural Networks." (2015).
Code: https://github.com/karpathy/char-rnn

Course 6.S094: Lex Fridman: Website: January

References: [35, 39] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Applications: Character-Level Text Generation
Life Is About The Weather!
Life Is About The (Wild) Truth About Human-Rights
Life Is About The True Love Of Mr. Mom
Life Is About Where He Were Now
Life Is About Kids
Life Is About What It Takes If Being On The Spot Is Tough
Life Is About An Eating Story
Life Is About The Truth Now

The meaning of life is literary recognition.

The meaning of life is the tradition of the ancient human reproduction

Andrej Karpathy. The Unreasonable Effectiveness of

Recurrent Neural Networks." (2015).
Code: https://github.com/karpathy/char-rnn

Course 6.S094: Lex Fridman: Website: January

References: [35, 39] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Applications: Image Caption Generation

Course 6.S094: Lex Fridman: Website: January

References: [43 Fang et al. 2015] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Applications: Image Question Answering

Ren et al. "Exploring models and data for image

question answering." 2015.
Code: https://github.com/renmengye/imageqa-public

Course 6.S094: Lex Fridman: Website: January

References: [40] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Applications: Video Description Generation

Venugopalan et al.
"Sequence to sequence-video to text." 2015.

Code: https://vsubhashini.github.io/s2vt.html

Course 6.S094: Lex Fridman: Website: January

References: [41, 42] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Applications: Modeling Attention Steering

Jimmy Ba, Volodymyr Mnih, and Koray

Kavukcuoglu. "Multiple object recognition
with visual attention." (2014).

Course 6.S094: Lex Fridman: Website: January

References: [35, 36] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Applications: Drawing with Selective Attention

Jimmy Ba, Volodymyr Mnih, and Koray

Kavukcuoglu. "Multiple object recognition
with visual attention." (2014).
Gregor et al. "DRAW: A recurrent neural network for image generation." (2015). Code: https://github.com/ericjang/draw

Course 6.S094: Lex Fridman: Website: January

References: [35, 37, 38] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Applications: Adding Audio to Silent Film

Owens, Andrew, Phillip Isola, Josh McDermott,

Antonio Torralba, Edward H. Adelson, and William T.
Freeman. "Visually Indicated Sounds." (2015).

Course 6.S094: Lex Fridman: Website: January

References: [28, 29] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Moravecs Paradox: The Easy Problems are Hard

Soccer is harder than Chess

Course 6.S094: Lex Fridman: Website: January

References: [8, 9] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Moravecs Paradox: The Easy Problems are Hard

Course 6.S094: Lex Fridman: Website: January

References: [8, 9] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Question: Why? Answer: Data

Visual perception: 540 millions years of data

Bipedal movement: 230+ million years of data
Abstract thought: 100 thousand years of data

Encoded in the large, highly evolved sensory and motor portions of the human brain is a
billion years of experience about the nature of the world and how to survive in it.
Abstract thought, though, is a new trick, perhaps less than 100 thousand years old. We have
not yet mastered it. It is not all that intrinsically difficult; it just seems so when we do it.
- Hans Moravec, Mind Children (1988)

Hans Moravec (CMU) Rodney Brooks (MIT) Marvin Minsky (MIT)

Course 6.S094: Lex Fridman: Website: January

References: [6, 7, 11] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Walking is Hard. How Hard is Driving?
Human performance: 1 fatality per 100,000,000 miles

Error rate for AI to improve on: 0.000001%

Challenges:
Snow
Heavy rain
Big open parking lots
Parking garages
Any pedestrian behaving irresponsibly or just unpredictably
Reflections, dynamics blinding ones
Merging into a high-speed stream of oncoming traffic
Course 6.S094: Lex Fridman: Website: January
References: [44] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Google Self-Driving Car: Driver Disengagements

Course 6.S094: Lex Fridman: Website: January

References: [77] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Google Self-Driving Car: Driver Disengagements

Course 6.S094: Lex Fridman: Website: January

References: [77] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Robustness:
>99.6% Confidence in the Wrong Answer

Nguyen et al. "Deep neural networks are easily fooled: High confidence predictions for unrecognizable images." 2015.
Course 6.S094: Lex Fridman: Website: January
References: [67] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Robustness:
Fooled by a Little Distortion

Szegedy et al. "Intriguing properties of neural networks." 2013.

Course 6.S094: Lex Fridman: Website: January
References: [68] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Mark I Perceptron
Frank Rosenblatt
400 pixel image input
Weights encoded in potentiometers
Weight updated by electric motors

NEW NAVY DEVICE LEARNS BY DOING

July 8, 1958
The Navy revealed the embryo of an electronic
computer today that it expects will be able to walk,
talk, see, write, reproduce itself and be conscious of
its existence Dr. Frank Rosenblatt, a research
psychologist at the Cornell Aeronautical Laboratory,
Buffalo, said Perceptrons might be fired to the
planets as mechanical space explorers

Course 6.S094: Lex Fridman: Website: January

References: [45] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Smaller episodes:
AI Winters
1966: the failure of machine translation
Two major episodes: 1970: the abandonment of connectionism
1974-80 1971-75: DARPA's frustration with the Speech
1987-93 Understanding Research program
1973: the large decrease in AI research in the UK
in response to the Lighthill report.
197374: DARPA's cutbacks to academic AI
research in general
1987: the collapse of the Lisp machine market
1988: the cancellation of new spending on AI by
the Strategic Computing Initiative
1993: expert systems slowly reaching the bottom
1990s: the quiet disappearance of the fifth-
generation computer project's original goals.

In no part of the field have discoveries made so far

produced the major impact that was then promised.
Course 6.S094: Lex Fridman: Website: January
References: [18] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 The Seasons of Deep Learning
1940s-1960s: Cybernetics
Biological learning (1943)
Perceptron (1958)
1980s-1990s: Connectionism
Backpropagation
2006-: Deep Learning

Course 6.S094: Lex Fridman: Website: January

References: [18] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 3rd Summer of Deep Learning

Google trends: Deep Learning

Course 6.S094: Lex Fridman: Website: January

References: [21] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Proceed with Caution:
Whats Next for Deep Learning?
(5 year vision)

Ilya Sutskever, Research Director of OpenAI:

Deeper models, models that need fewer examples for training.
Christian Szegedy, Senior Research Scientist at Google:
Become so efficient that they will be able to run on cheap mobile devices.
Pieter Abbeel, Associate Professor in Computer Science at UC Berkeley:
Significant advances in deep unsupervised learning and deep reinforcement
learning.
Ian Goodfellow, Senior Research Scientist at Google:
Neural networks that can summarize what happens in a video clip, and will be
able to generate short videos. Neural networks that model the behavior of
genes, drugs, and proteins and then used to design new medicines.
Koray Kavukcuoglu & Alex Graves, Research Scientists at Google DeepMind:
An increase in multimodal learning, and a stronger focus on learning that
persists beyond individual datasets.
Charlie Tang, Machine Learning group, University of Toronto:
Deep learning algorithms ported to commercial products, much like how the
face detector was incorporated into consumer cameras in the past 10 years.

Course 6.S094: Lex Fridman: Website: January

References: [58] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Gartner Hype Cycle

Course 6.S094: Lex Fridman: Website: January

References: [21] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Course 6.S094: Lex Fridman: Website: January
References: [51] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Course 6.S094: Lex Fridman: Website: January
References: [51] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
Attention to (AI) Drivers: Proceed with Caution

Camera Spoofing LIDAR Spoofing

Course 6.S094: Lex Fridman: Website: January

References: [68, 72] Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Interface: Python, (C++)
Automatic Differentiation
Multi GPU, Cluster Support
Currently most popular

Course 6.S094: Lex Fridman: Website: January

Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Keras

On top of Tensorflow (and Theano)

Interface: Python
Goal: provide a simplified interface

Also: TF Learn, TF Slim

Course 6.S094: Lex Fridman: Website: January

Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Torch

Used by researchers doing lower level (closer to the details)

neural net work
Interface: Lua
Fragmented across different plugins

Course 6.S094: Lex Fridman: Website: January

Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Interface: Python (tight NumPy integration)
One of the earlier frameworks with GPU support
Encourages low-level tinkering

Course 6.S094: Lex Fridman: Website: January

Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 cuDNN

The library that most frameworks use for doing the actual
computation
Implements primitive neural network functions in CUDA on
the GPU

Course 6.S094: Lex Fridman: Website: January

Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Multi GPU Support (scales well)
Interface: Python, R, Julia, Scala, Go, Javascript

Course 6.S094: Lex Fridman: Website: January

Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Interface: Python
Often best on benchmarks
Nervana was working on a neural network chip
Bought by Intel

Course 6.S094: Lex Fridman: Website: January

Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Caffe

Interface: C++, Python

One of the earliest GPU supported
Initial focus on computer vision (and CNNs)

Course 6.S094: Lex Fridman: Website: January

Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 Microsoft Cognitive Toolkit (CNTK)

Interface: Custom Language (BrainScript), Python, C++, C#

Multi GPU Support (scales very well)
Mostly used at MS Research

Course 6.S094: Lex Fridman: Website: January

Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 In the Browser

Keras.js
GPU Support
Full sized networks
Can use trained Keras models

ConvNetJS
Built by a Andrej Karpathy
Good for explaining neural network concepts
Fun to play around with
Very few requirements
Full CNN, RNN, Deep Q Learning

Course 6.S094: Lex Fridman: Website: January

Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017
 References

All references cited in this presentation are listed in the

following Google Sheets file:

https://goo.gl/9Xhp2t

Course 6.S094: Lex Fridman: Website: January

Deep Learning for Self-Driving Cars [email protected] cars.mit.edu 2017