Beruflich Dokumente
Kultur Dokumente
Introduction
Deep learning (also known as deep structured learning or hierarchical
learning) is part of a broader family of machine learning methods based
on learning data representations, as opposed to task-specific algorithms.
Learning can be supervised, semi-supervisedor unsupervised.
Extract
Gather Clean
Featur Learn Predict
Data Data
es
2
Feature Extraction
b3
Hidden 2: 100
X W1 W2 W3
Softmax
ReLU
ReLU
y
Output: 10 classes
Hidden 1: 300
b2
• ReLU: 𝑓 𝒙 = max 0, 𝒙
𝑒 𝑥𝑖
• Softmax: 𝑓 𝒙 = σ𝑛 𝑥
𝑖=1 𝑒 𝑖
• Cross Entropy: 𝑓 𝒂, 𝒃 = − σ 𝒂 ∙ log 𝒃
• SGD: 𝒘′ = 𝒘 − 𝑙 ∙ ∆𝒘, ∀𝒘
Convolution Neural Networks (CNN)
IBM RESEARCH
Recurrent Neural
Network
.
• Data is temporal/seqential
RNN – Encoder Decoder Model
Language A Language B
Question Answer
Sentence Encode Decode Paraphrase
Compound word r r Constituent
Paragraph words
... Summary
...
Sequence-to-Sequence models
IBM RESEARCH
Generative Adversarial Networks
• Generative Network:
• Producing images that are closer to real images.
• Discriminative Network:
• Determine difference between real and fake images.
Aim: Generative network that can produce images which are indistinguishable
from the real ones.
Use cases:
• Generating realistic artwork samples (video/image/audio)
• Image to Image Translation
• Image super resolution
• Image Inpainting
• Text to Image
IBM RESEARCH
Generative Adversarial Networks
Deep Learning Libraries
Three Competitive Advantages of Deep
Learning for the Company
1. Cost and time benefits.
Neural networks are trainable “brains.” You give them your company’s
information and train them to do a task, such as generating reports, and
they will use that training, new information, and their “work experience” to
adapt and improve in much the same way a human worker learns.