Building products with

ML & TensorFlow

@svonava, svonava@gmail.com, svonava.com

How is ML used today (MIT Tech Review*, 2016)

Already achieved

Objectives

*375 C-suite/VP/enterprise devs across USA, EU & India

How is ML used today (MIT Tech Review, 2016)

Primary driver for ML?

How is ML used today (MIT Tech Review, 2016)

% of 2017 budget dedicated to ML

Strategy & tactics

But first a bit of history!

“a machine which senses, recognizes, remembers, and responds like the human mind”

The New Yorker

But first a bit of history!

“a machine which senses, recognizes, remembers, and responds like the human mind”

The New Yorker, 1958

The 60ties

Rosenblatt

The 60ties

playground.tensorflow.com

The 80ties

playground.tensorflow.com

The 80ties

playground.tensorflow.com

Feature engineering!

Back propagation!

Deep[er] learning

playground.tensorflow.com

Deep learning (mNeuron)

Deep learning (mNeuron)

Convolution!

Remember late 2010 & mobile first?

StatCounter

Strategy & tactics

Strategy: Talent

Hire

• don’t ask a Quant to figure out your business model
• design autonomy
• $$$ for compute & data acquisition
• Never alone!

Grow

• ML ninja (SWE + PM)
• project incubation
• ML office hours/consulting

Spread!

ML guidelines, internal training, open-sourcing

Strategy: Infrastructure

Pre-trained APIs → Custom Models

General comp. (GPU) → ASIC (TPU)

​

​

Pipelines (pre/post processing)

Efficient model training (ASIC)

Model serving

or deploy to mobile

Interactive analysis

Data acquisition

Storage

Pre-trained model APIs

Strategy: Infrastructure

Strategy: Frameworks

Static comp. graph�Tensorflow

​

Tooling, deployment, domain-specific abstractions ...

Dynamic comp. graph�PyTorch

​

Variable length data is padded by zeros etc.�→ Batching & efficiency

Forward comp. like numpy�→ Easier adoption

i = tf.constant(0)�c = lambda i: tf.less(i, 10)�b = lambda i: tf.add(i, 1)�r = tf.while_loop(c, b, [i])

​

for _ in range(T):� h = torch.matmul(W, h) + b

Strategy: Risk

What can go wrong?

Strategy: Risk

Strategy: Risk

• Self-driving autopilot failures.
• Amplification of training-data biases saw failures in Insurance, Law enforcement, Health care.
• Skin-color & variation in shapes of human body cause AI Vision failures.
• User-generated content & advertising.

​

→ Invest in human oversight, fallback solutions & PR/expectation management.

Strategy: Ethics

Disruption of worker & employer negotiation, hiding behind ‘neutral’ technology.

Simple

• Drivers don’t see the destination, so they don’t cherry-pick profitable jobs.

ML

• Drivers are manipulated through notifications to be available during off-peak.
• Social networks allow targeting such as ‘insecure teens’.
• Elections & public manipulation using micro-targeting.

→ No standard ethics code and no way to transparently enforce… but check AI Now.

Strategy: Ethics

60% of all jobs have 30% of tasks readily automatable (McKinsey).

Strategy & tactics

Personalization

Automation

Anticipation

Forecasting

Data processing

Anomaly detection

external

internal

Personalization

Automation

Anticipation

Forecasting

Data processing

Anomaly detection

external

internal

UX for window-shopper vs fire-and-forget.

Receipt input for expenses.

Reactive → Proactive (Search vs. Google Now)

Web traffic forecast for Ad selling.

Sentiment analysis on users to prioritize dev.

Fraud prevention.

1. What problem are we solving?

• New feature vs. more automatic/faster?
• Do the users want it? How much value? User journeys?
• Success metrics?
• Quality, latency, what do we measure against?
• Risks?
• How can the solution backfire? How much damage? Adversarial tests?

​

2. What data do we need to collect?

• What features:
• as fine-grained as possible (notes vs live perf story)
• don’t stress ‘feature engineering’ (evolution of vision features story)
• IDs as features, the best for personalization but sucks for forecasting/transfer learning (need user clusters for that)
• Needs to capture 2 spaces:
• Feature-to-feature - “What videos are watched in session by a person?”
• Feature-to-label - “People that watched X are more passive.”
• If we need labels, how do we get them?
• Ask users? Human annotation? Expert-based rules?

​

• And how much data?

Degrees of freedom:

• 1000s of examples of each class
• 10s of examples of every feature
• 10s of examples for every model parameter

​

Complexity of the model:

• Simple linear: 100s
• Shallow NN: 100,000s
• Deep learning: Tens of millions or more

​

→ More data & simpler models.

→ The less independent the examples, the more of them you need.

→ Easy test: How much can you sample/downscale the examples → gives “useful bits”.

3. Executing the ML solution

• Pick algorithms based on the success metrics & the data we could get.
• Pick a simple one and invest 50% time into building quality eval (ask users if helpful, track their behavior changes, backtest forecasts, human evals etc).
• Build an experiment framework (exposed vs control) for eval & release process.�→ Feedback loops!

Before AI, go HI

1. Test the product idea using a ‘human intelligence’ backend.
2. Scale up using mechanical turk APIs.�→ Hello training data!
3. Roll-out ML.

Before AI, go HI

Google Surveys

• targeting: android users, zip-code, screening questions
• can be multi-question to identify relationships
• $0.3/response for single-question
• $3/response for multi-question

​

Before AI, go HI

body_def = {� 'title': 'Student cell phone ownership',� 'description': 'Ownership of cell.',� 'owners': owner_emails,� 'wantedResponseCount': 100,� 'audience': {� 'country': 'US',� },� 'questions': [� {� 'question': 'Do you own a cell phone?',� 'type': 'singleAnswer',� 'answers': ['Yes', 'No'],� }� ]�}�surveys_service.surveys().insert(body=body_def).execute()

​

Before AI, go HI

Amazon Mechanical Turk

• More complex tasks (image annotation, extract data from websites etc).
• Targeting: demo, country, language, but also stuff like ‘holds ETFs’ or ‘exercise 2-3 times a week’, etc.
• Pricing: Worker reward up to you ($0.05/simple task at least).
• See API sample - still pretty easy.

→ Near-real time (x minutes) with:

Alegion, StintMint, SquadRun (also possible with Amazon)

​

Getting labeled data for free?

Check out human-based computation games (wikipedia).

Scaling & latency beyond HI - pre-trained models!

Case study: Natural Language Processing (Wootric)

TensorFlow: Tensor-glove (Google Cloud) - low entry bar

Input: Couple 100s labeled points (3 classes) in 3D.

TensorFlow: Tensor-glove (Google Cloud)

​

​

Steps:

1. Visualize & notice the linear separability.

TensorFlow: Tensor-glove (Google Cloud)

Steps:

• Visualize & notice the linear separability.
• Train a model y_classes = W_weights * x_inputs + b_biases

TensorFlow: Tensor-glove (Google Cloud)

# define the model�import tensorflow as tf�weights = tf.Variable(tf.zeros([3, 3]), dtype=tf.float32) # The model learns this�biases = tf.Variable(tf.zeros([3]), dtype=tf.float32) # and this.�glove_data = tf.placeholder(tf.float32) # Sensor data goes here e.g. [[5 .3 47]..]. �rps_data = tf.matmul(glove_data, weights) + biases # Encodes y=Wx+b.��# define the loss�rps_labels = tf.placeholder(tf.float32) # Labels for glove_data e.g. [[1 0 0]..].

loss = tf.losses.softmax_cross_entropy(onehot_labels=rps_labels, logits=rps_data)

​

​

Softmax - squish rps_data to [0, 1] with sum = 1 (probabilities)

Cross entropy - diff of the probabilities and the one-hot representation

TensorFlow: Tensor-glove (Google Cloud)

# define an optimizer�optimizer = tf.train.GradientDescentOptimizer(learning_rate=1).minimize(loss)�# create a Session�sess = tf.Session()�# init the variables�initializer = tf.global_variables_initializer()�sess.run(initializer)�# train the model�for i in range(10000):� # train the model� sess.run(optimizer, {glove_data: g_data, rps_labels: g_label})� # print current loss� if i % 1000 == 0:� l = sess.run(loss, {glove_data: g_data, rps_labels: g_label})

​

TensorFlow: Tensor-glove (Google Cloud)

What’s next?

• Generative adversarial networks (GANs) (log forecasting story)
• Variational autoencoders (VANs)

​

Thank you!

Follow @svonava for more generative goodness coming soon!