artificial intelligence (a.i.), machine and deep learning...

30
Artificial Intelligence (A.I.), Machine and Deep Learning past, present and future aspects Bálint Gyires-Tóth

Upload: others

Post on 26-Sep-2020

6 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

Artificial Intelligence (A.I.), Machine and Deep Learningpast, present and future aspectsBálint Gyires-Tóth

Page 2: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

About me

13/06/2018Bálint Gyires-Tóth, PhD ([email protected])

Bálint Gyires-Tóth, PhD

SmartLab @ BME TMIT

NVidia Deep Learning Institute Certified Instructor & University Ambassador

- Signal Processing

- Time series modeling

- Machine & deep learning since 2007

- Research & Development

- Education & Trainings

2/30

Page 3: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

Past…

13/06/2018Bálint Gyires-Tóth, PhD ([email protected]) 3/30

Page 4: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

13/06/2018Bálint Gyires-Tóth, PhD ([email protected]) 4/30

Page 5: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

13/06/2018Bálint Gyires-Tóth, PhD ([email protected]) 5/30

Page 6: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

AI: 4th generation of knowledge acquisition

1. Evolution

2. Tacit knowledge

3. Explicit knowledge

4. AI

13/06/2018Bálint Gyires-Tóth, PhD ([email protected]) 6/30

Page 7: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

A.I., Machine Learning, Deep Learning

13/06/2018Bálint Gyires-Tóth, PhD ([email protected]) 7/30

Page 8: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

Challenge

Analitical and data driven thinking

Data driven thinking: Maslow’s hammer

13/06/2018Bálint Gyires-Tóth, PhD ([email protected]) 8/30

Page 9: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

Present…

13/06/2018Bálint Gyires-Tóth, PhD ([email protected]) 9/30

Page 10: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

2014 – Deep Mind, ~500.000.000$ (https://deepmind.com)• Space Invaders

• Breakout

• AlphaGo

• AlphaGo Zero

2015 – OpenAI, ~1.000.000.000$ (https://openai.com)• Elon Musk (PayPal, SpaceX, Tesla Motors) + investors• Open source AI soulitions for mankind

2016 – Nervana – Intel fusion

2017 – Maluuba (Montreal, Canada) – Microsoft fusion

2018 – France – 1.850.000.000$ AI Research

To be continued…13/06/2018Bálint Gyires-Tóth, PhD ([email protected]) 10/30

Page 11: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

Deep learning based AI is the next big possibility for growth!

I works! Superhuman level…

~ Hearing ~ Vision

~ Speech ~ Knowledge?

13/06/2018Bálint Gyires-Tóth, PhD ([email protected]) 11/30

Page 12: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

The Holy Grail of data science is Deep Learning?

‘Deep learning is part of a broader family of machine learning methods based on learning data representations.’

In practice: deep neural networks

Whats new?

Algorithms + data + GPU

+ lower entry level

+ open-source research society (‘democratizing AI’)

Bálint Gyires-Tóth, PhD ([email protected]) 13/06/2018 12/30

Page 13: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

Feed forward neural networks

13/06/2018

f(∑): nonlinearn functions

e.g. f(∑)=max(0, ∑)

backpropagation

LEARNING: „fine-tuneing” weights

Bálint Gyires-Tóth, PhD ([email protected]) 13/30

Page 14: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

Basic layer types

• Fully Connected layers, FC• Classification and regression

• Recurrent layers (e.g. Long Short-Term Memory, LSTM)• Sequential and temporal data

• Convolutional layers (Convolutional Neural Net, CNN)• Feature extraction vs feature (representation) learning

• 1D, 2D and 3D convolution

• Initially for images and speech, nowadays for general purposes

13/06/2018Bálint Gyires-Tóth, PhD ([email protected]) 14/30

Page 15: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

Hardver: GPU

NVidia GTX, Titan, Titan X, Xp, V• Titan X Pascal: 12 teraFlop (single precision)

• ~500..2800 EUR

NVidia DevBox: 40-50 TFlop, 15k EUR

NVidia DGX Station: 480 TFlop, 65k EUR

NVidia DGX-1: 960 TFlop, 220k EUR

Nvidia DGX-2: 2 PFlop (2000 TFlop), 560k EUR

3rd party vendors

Cloud services

13/06/2018Bálint Gyires-Tóth, PhD ([email protected]) 15/30

Page 16: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

SoftwareTensorFlow (Google)

Keras (Fchollet @Google)

PyTorch, Torch7 (Facebook, Twitter, NVidia, stb.)

Theano (University of Montreal)

Nervana (Intel), Lasagne, Microsoft CNTK, Matlab Deep Learning Toolbox, stb

Caffe (University of California, Berkeley)

Kaldi

13/06/2018Bálint Gyires-Tóth, PhD ([email protected]) 16/30

Page 17: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

Deep Learning exampleImage recognition

13/06/2018 Bálint Gyires-Tóth, PhD ([email protected]) 17

Page 18: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

ImageNet challenge

• ~1.200.000 images, ~1000 categories

• Goal: Top-5 accuracy

• AlexNet [A. Krizhevsky et al., 2012]

13/06/2018Bálint Gyires-Tóth, PhD ([email protected]) 18/30

Page 19: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

AlexNet [A. Krizhevsky et al., 2012]

13/06/2018Bálint Gyires-Tóth, PhD ([email protected]) 19/30

Page 20: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

Top-5 accuracy 2010-2015

13/06/2018

https://devblogs.nvidia.com/parallelforall/mocha-jl-deep-learning-julia/

Bálint Gyires-Tóth, PhD ([email protected]) 20/30

Page 21: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

from urllib.request import urlretrieve

from keras.applications.inception_v3 import InceptionV3,preprocess_input,decode_predictions

from keras.preprocessing import image

import numpy as np

url_dog="https://pixabay.com/static/uploads/photo/2016/02/19/15/46/dog-1210559_960_720.jpg"

urlretrieve(url_dog, "dog.jpg")

img_path = 'dog.jpg'

img = image.load_img(img_path, target_size=(299, 299))

x = image.img_to_array(img)

x = np.expand_dims(x, axis=0)

x = preprocess_input(x)

model = InceptionV3(weights='imagenet', include_top=True)

preds = model.predict(x)

print('Predicted class:', decode_predictions(preds))

13/06/2018Bálint Gyires-Tóth, PhD ([email protected]) 21

Page 22: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

from urllib.request import urlretrieve

from keras.applications.inception_v3 import InceptionV3,preprocess_input,decode_predictions

from keras.preprocessing import image

import numpy as np

url_dog="https://pixabay.com/static/uploads/photo/2016/02/19/15/46/dog-1210559_960_720.jpg"

urlretrieve(url_dog, "dog.jpg")

img_path = 'dog.jpg'

img = image.load_img(img_path, target_size=(299, 299))

x = image.img_to_array(img)

x = np.expand_dims(x, axis=0)

x = preprocess_input(x)

model = InceptionV3(weights='imagenet', include_top=True)

preds = model.predict(x)

print(‘Predicted class:', decode_predictions(preds))

13/06/2018Bálint Gyires-Tóth, PhD ([email protected]) 22

Page 23: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

from urllib.request import urlretrieve

from keras.applications.inception_v3 import InceptionV3,preprocess_input,decode_predictions

from keras.preprocessing import image

import numpy as np

url_dog="https://pixabay.com/static/uploads/photo/2016/02/19/15/46/dog-1210559_960_720.jpg"

urlretrieve(url_dog, "dog.jpg")

img_path = 'dog.jpg'

img = image.load_img(img_path, target_size=(299, 299))

x = image.img_to_array(img)

x = np.expand_dims(x, axis=0)

x = preprocess_input(x)

model = InceptionV3(weights='imagenet', include_top=True)

preds = model.predict(x)

print('Predicted class:', decode_predictions(preds))

13/06/2018Bálint Gyires-Tóth, PhD ([email protected])

Predictions:[[('n02099712', 'Labrador_retriever', 0.5143939), ('n02099601', 'golden_retriever', 0.28467834), ('n02108551', 'tennis_ball ', 0.006718181), ('n02104029', Rhodesian_ridgeback', 0.0056073326), ('n04409515', kuvasz', 0.0041805333)]]

23

Page 24: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

Deep learning applications

• Images• Classification, object detection, segmentation

• Neural art, style transfer

• Sequential and temporal data• Speech recognition and synthesis

• Natural Language Processing (eg. text, intent classification)

• Telecommunication data (logs, traffic, anomalies, etc.)

• Medical data (EEG, ECG signals)

• Financial data

• Corporate Data• Customer analysis – e.g. fraud, churn prediction

• Risk modeling

• Marketing

13/06/2018Bálint Gyires-Tóth, PhD ([email protected]) 24/30

Page 25: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

Some interestingDeep learning applications

13/06/2018 Bálint Gyires-Tóth, PhD ([email protected]) 25

Page 26: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

WaveNet: SOTA Text-To-Speech Engine

Inputs:

4k-8k previous samples (0.2-0.5 s) +

contextual features

Layers:

40-80 dilated convolutional layers

256 neurons in an FC layer

Output:

256 level μ-law quantized

raw waveform

(based on DeepMind’s WaveNet)

13/06/2018Bálint Gyires-Tóth, PhD ([email protected])

inputs

output

1d causal

CONV

1d dilated

CONV

1d 1x1

CONV

1d 1x1

CONV

1d 1x1

CONV

#trainable parameters: 1.2 million

26/30

Page 27: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

Modeling sensor data

Sensors• Gyroscope

• Orientation

• GPS, WiFi, etc.

Deep learning based classification• Activity

• Behaviour

• Body gestures

Applications• Emergency

• Activity analysis

• User and user group identification

13/06/2018Bálint Gyires-Tóth, PhD ([email protected]) 27/30

Page 28: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

Sensor data: best model

13/06/2018Bálint Gyires-Tóth, PhD ([email protected])

INPUT:

200 x 7 sample points (1-2 seconds)

LAYERS (HYPEROPT):

16 residual blocks

filter (16+) and kernel size (3,5,7,9,11,…)

OUTPUT:

Binary classification

Over 99% accuracy.

28/30

Page 29: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

(Near) future…

High level frameworks.

Standardized workflows and model formats.

Easier and faster deployment.

More corporate / industrial applications.

Better tools for Deep Reinforcement Learning.

13/06/2018Bálint Gyires-Tóth, PhD ([email protected]) 29/30

Page 30: Artificial Intelligence (A.I.), Machine and Deep Learning ...biconsulting.hu/letoltes/...balint_deep_learning_past_present_future.pdf · (Near) future… High level frameworks. Standardized

Thank you!Bálint Gyires-Tóth

[email protected]