These slides accompanied a demo of Deeplearning4j at the SF Data Mining Meetup hosted by Trulia. http://www.meetup.com/Data-Mining/events/212445872/ Deep-learning is useful in detecting identifying similarities to augment search and text analytics; predicting customer lifetime value and churn; and recognizing faces and voices. Deeplearning4j is an infinitely scalable deep-learning architecture suitable for Hadoop and other big-data structures. It includes a distributed deep-learning framework and a normal deep-learning framework; i.e. it runs on a single thread as well. Training takes place in the cluster, which means it can process massive amounts of data. Nets are trained in parallel via iterative reduce, and they are equally compatible with Java, Scala and Clojure. The distributed deep-learning framework is made for data input and neural net training at scale, and its output should be highly accurate predictive models. The framework's neural nets include restricted Boltzmann machines, deep-belief networks, deep autoencoders, convolutional nets and recursive neural tensor networks. Finally, Deeplearning4j integrates with GPUs. A stable version was released in October.

1. Deep Learning
{
Machine Perception and Its Applications
Adam Gibson // deeplearning4j.org // skymind.io // zipfian

2. DL, a Subset of AI
Deep Learning = subset of Machine Learning
Machine Learning = subset of AI
 AI = Algorithms that repeatedly optimize
themselves.
 Deep learning = pattern recognition
Machines classify data and improve over
time.

3. Why Is DL Hard?
We see this… Machines see this… (Where’s the cat?)
(Hat tip to Andrew
Ng)

4. What Can It Handle?
Anything digitized
 Raw media: MP3’s, JPEG’s, text, video
 Sensor output: temperature, pressure,
motion and chemical composition
Time-series data: Prices and their
movement; e.g. the stock market, real
estate, weather and economic indicators
 It’s setting new accuracy records
everywhere

5. What’s It Good For?
Recommendation engines: Anticipate what
you will buy or click.
 Anomaly detection: Bad outcomes signal
themselves in advance: fraud in e-commerce;
tumors in X-rays; loans likely to
default.
Signal processing: Deep learning can
estimate customer lifetime value, necessary
inventory or an approaching market crash.
Facial and image recognition

6. Facial
recognition
& feature
hierarchy
(Hat tip to Andrew
Ng)

7. DL4J Facial Reconstructions

8. How Did It Do That?
Nets need training data.
You know what training sets contain.
Nets learn training-set faces by repeated
reconstruction.
Reconstruction = finding which facial
features are indicative of larger forms.
When a net can rebuild the training set, it is
ready to work with unsupervised data.

9. Technical Explanation
Nets measure the difference between their
results and a benchmark = loss function
They minimize differences with an
optimization function.
They optimize by altering their parameters
and testing how changes affect results.
Gradient descent, Conjugate gradient, L-BFGS

10. Learning looks like this.
Note the local minima…

11. Representation Learning
Through pre-training, nets learn to locate
signal in a world of noise
Generic priors initiate weights
Reconstructions = representations
Feature hierarchies  intuition about
complex, abstract features

12. Facial Recognition’s Uses
Facebook engages us more. (95-97%
accuracy)
Government agencies identify persons of
interest.
Video game makers build more realistic
(and stickier) worlds.
 Stores identify customers and track
behavior, prevent churn and encourage
spending.

13. Sentiment Analysis & Text
 Sentiment analysis ~ NLP
Software classifies sentences by emotional
tone, bias and intensity
Positive or negative - object-specific…
Rank movies, books, consumer goods,
politicians, celebrities
 Predict social unrest, gauge reputations, PR…

14. Restricted Boltzmann
Machine (RBMs)

15. Deep-Belief Net (DBN)
A stack of RBMs.
1st RBM’s hidden layer -> 2nd RBM’s input layer
Feature hierarchy
A DBN classifies data.
Buckets images: e.g. sunset, elephant, flower.
Useful in search.

16. Deep Autoencoder
Two DBNs.
The first DBN encodes data into vector of
10-30 numbers.
 The second DBN decodes data back to
original state.
Reduce any document/image to highly
compact vector.
QA and information retrieval: Watson

17. Some Results

18. Image Search Results

19. Convolutional Net
Good with images.
ConvNets learn data like images in
patches.
 Each piece learned is then woven
together in the whole.
Yann LeCun’s baby, now at Facebook.

20. Recursive Neural Tensor Net
Top-down, hierarchical nets rather than feed-forward
like DBNs.
Sequence-based classification, windows of
several events, entire scenes rather than
images.
Features = vectors.
A tensor = multi-dimensional matrix, or
multiple matrices of the same size.

21. RNTNs & Scene Composition

22. RNTNs & Sentence Parsing

23. t-SNE for Data Visualizations

24. DL4J + MNIST + t-SNE