Classification

With good data, deep learning can give high accuracy in image and text classification

It's trivially easy to train your own classifier
with near zero ML knowledge

It's so easy that ...

6th and 7th grade high school students created a custom vision classifier for TrashCam

[Trash, Recycle, Compost] with 90% accuracy

Intracranial Hemorrhage

Work by MM colleagues: Caiming Xiong, Kai Sheng Tai, Ivo Mihov, ...

Advances leveraged via GPUs

AlexNet training throughput based on 20 iterations

Slide from Julie Bernauer's NVIDIA presentation

Beyond classification ...

VQA dataset: http://visualqa.org/

Beyond classification ...

* TIL Lassi = popular, traditional, yogurt based drink from the Indian Subcontinent

Question Answering

Visual Genome: http://visualgenome.org/

Question Answering

Visual Genome: http://visualgenome.org/

Question Answering

1 Mary moved to the bathroom.
2 John went to the hallway.
3 Where is Mary?        bathroom        1
4 Daniel went back to the hallway.
5 Sandra moved to the garden.
6 Where is Daniel?      hallway         4
7 John moved to the office.
8 Sandra journeyed to the bathroom.
9 Where is Daniel?      hallway         4
10 Mary moved to the hallway.
11 Daniel travelled to the office.
12 Where is Daniel?     office          11
13 John went back to the garden.
14 John moved to the bedroom.
15 Where is Sandra?     bathroom        8
1 Sandra travelled to the office.
2 Sandra went to the bathroom.
3 Where is Sandra?      bathroom        2

Extract from the Facebook bAbI Dataset

Human Question Answering

Imagine I gave you an article or an image, asked you to memorize it, took it away, then asked you various questions.

Even as intelligent as you are,
you're going to get a failing grade :(

Why?

You can't store everything in working memory
Without a question to direct your attention,
you waste focus on unimportant details

Optimal: give you the input data, give you the question, allow as many glances as possible

Think in terms of

Information Bottlenecks

Where is your model forced to use a compressed representation?

Most importantly,
is that a good thing?

Gated Recurrent Unit (GRU)

Cho et al. 2014

A type of recurrent neural network (RNN), similar to the LSTM
- Consumes and/or generates sequences (chars, words, ...)
The GRU updates an internal state h according to the:
- existing state h and the current input x

h_t = GRU(x_t, h_{t-1})

h_t = GRU(x_t, h_{t-1})

Figure from Chris Olah's Visualizing Representations

Neural Machine Translation

Figure from Chris Olah's Visualizing Representations

Figure from Bahdanau et al's
Neural Machine Translation by Jointly Learning to Align and Translate

Neural Machine Translation

Results from Bahdanau et al's
Neural Machine Translation by Jointly Learning to Align and Translate

Related Attention/Memory Work

Sequence to Sequence (Sutskever et al. 2014)
Neural Turing Machines (Graves et al. 2014)
Teaching Machines to Read and Comprehend
(Hermann et al. 2015)
Learning to Transduce with Unbounded Memory (Grefenstette 2015)
Structured Memory for Neural Turing Machines
(Wei Zhang 2015)
Memory Networks (Weston et al. 2015)
End to end memory networks (Sukhbaatar et al. 2015)

QA for Dynamic Memory Networks

A modular and flexible DL framework for QA
Capable of tackling wide range of tasks and input formats
Can even been used for general NLP tasks (i.e. non QA)
(PoS, NER, sentiment, translation, ...)

Ask Me Anything: Dynamic Memory Networks for Natural Language Processing (Kumar et al., 2015)

Dynamic Memory Networks for Visual and Textual Question Answering (Xiong et al., 2016)

For full details:

QA for Dynamic Memory Networks

A modular and flexible DL framework for QA
Capable of tackling wide range of tasks and input formats
Can even been used for general NLP tasks (i.e. non QA)
(PoS, NER, sentiment, translation, ...)

Input Modules

+ The module produces an ordered list of facts from the input
+ We can increase the number or dimensionality of these facts
+ Input fusion layer (bidirectional GRU) injects positional information and allows interactions between facts

Episodic Memory Module

Composed of three parts with potentially multiple passes:

Computing attention gates
Attention mechanism
Memory update

Computing Attention Gates

Each fact receives an attention gate value from [0, 1]

The value is produced by analyzing [fact, query, episode memory]

Optionally enforce sparsity by using softmax over attention values

Soft Attention Mechanism

c = \sum^N_{i=1} g_i f_i

c = \sum^N_{i=1} g_i f_i

If the gate values were passed through softmax,
the context vector is a weighted summation of the input facts

Given the attention gates, we now want to extract a context vector from the input facts

Issue: summation loses positional and ordering information

Attention GRU Mechanism

If we modify the GRU, we can inject information from the attention gates.

By replacing the update gate u with the activation gate g,
the update gate can make use of the question and memory

Attention GRU Mechanism

If we modify the GRU, we can inject information from the attention gates.

For training,

GPUs are leading the way

VisualQA dataset has over 200k images and 600k questions

GPUs are the key to efficient training, especially at higher resolutions

The DMN make heavy use of RNNs

CNNs have experienced majority of optimization focus
(many optimizations are trivial)
RNNs on GPUs still have room to improve
NVIDIA are actively improving RNN optimization