SlideShare a Scribd company logo
1 of 40
Tsvi Lev,
Corporate VP and Head of NEC Labs Israel
tsvi.lev@necam.com
XAI Beyond the Hype
#MAXA
2 © NEC Corporation 2015 NEC Group Internal Use Only2 © NEC Corporation 2019
NEC – we too were once a startup..
3 © NEC Corporation 2015 NEC Group Internal Use Only3 © NEC Corporation 2019
120 Years of Research Innovation
4 © NEC Corporation 2015 NEC Group Internal Use Only4 © NEC Corporation 2019
Applied AI is quite different from Kaggle competitions or papers
NEC Labs do Applied AI for vision, healthcare, cyber, edge. Our Customers:
▌don’t care about SOTA or AUC
▌lack training data
▌want AI to behave ‘like us’
ACM CCS Best Paper Edge AI for Vision Cyber Research AI for Healthcare
5 © NEC Corporation 2015 NEC Group Internal Use Only5 © NEC Corporation 2019
▌ The recent AI craze started with Digital Marketing success ($$$) and
Academic breakthroughs in Image and Speech Processing
▌Other application areas have different requirements
The truth about recent AI success
More $$$!
Better than Human
6 © NEC Corporation 2015 NEC Group Internal Use Only6 © NEC Corporation 2019
▌ Recommendation Engines replaced rule based systems
▌ Measured only by results (clicks, sales)
▌ Recently – issues of Bias and Accountability
Use Case: digital marketing
Accuracy
Stability
Compliance
Explainablity
Accountability
Bias Free
7 © NEC Corporation 2015 NEC Group Internal Use Only7 © NEC Corporation 2019
▌Replace non-expert Humans
▌Acceptance threshold prevents ‘just better than previous’ criteria
▌Stability and Compliance are key from day 0
▌Hence – slower than promised and hyped
Use Case: Mobility
Accuracy
Stability
Compliance
Explainablity
Accountability
Bias Free
8 © NEC Corporation 2015 NEC Group Internal Use Only8 © NEC Corporation 2019
▌ Assist, not replace Expert Human Judgement.
▌ Xplain-ability and Stability are >> than accuracy
▌ Hence – slower, but progress is stable
Use Case : Medical Diagnostics
Accuracy
Stability
Compliance
Explainablity
Accountability
Bias Free
9 © NEC Corporation 2015 NEC Group Internal Use Only9 © NEC Corporation 2019
The Future will be dominated by XAI
“Confidence in AI only comes with explainablity and
transparency” – Well Known B2B AI vendor
10 © NEC Corporation 2015 NEC Group Internal Use Only10 © NEC Corporation 2019
When do we most need XAI the most?
Recommendation Engines Man Machine Interfaces
Online Advertising Gaming
Single Decision=> Lives, much $$$ Many Micro Decision Aggregated
11 © NEC Corporation 2015 NEC Group Internal Use Only11 © NEC Corporation 2019
What may happen without XAI – AUC/ROC deterioration
Populations, physicians, imaging machines change – how to ensure AI is constant?
Benign Malignant! Benign
12 © NEC Corporation 2015 NEC Group Internal Use Only12 © NEC Corporation 2019
Users attach different gravity to certain error cases
▌ Errors in “simple” cases: blood vessel vs nodules
▌ Distribution of errors: no false alarms, then suddenly a ton for no reason
▌ “Butterfly in Asia” – tiny changes in input yield large changes in the result
13 © NEC Corporation 2015 NEC Group Internal Use Only13 © NEC Corporation 2019
Which Models are a best fit for XAI?
14 © NEC Corporation 2015 NEC Group Internal Use Only14 © NEC Corporation 2019
Model type does not guarantee explainability
MNIST decision tree (Hinton et al)
15 © NEC Corporation 2015 NEC Group Internal Use Only15 © NEC Corporation 2019
Explainable can mean different things
Causality Analysis
Reliability/Abstention
Model Improvement
Pattern Discovery
Post-hoc: methods to be
applied to other models.
IBM Research
16 © NEC Corporation 2015 NEC Group Internal Use Only16 © NEC Corporation 2019
Causality Analysis
Local Interpretable Model-agnostic Explanations (LIME)
17 © NEC Corporation 2015 NEC Group Internal Use Only17 © NEC Corporation 2019
Reliability/Abstention Criteria
▌Threshold – estimate the reliability by the output magnitude
▌Distance to labeled examples –
⚫Let V be the output vector of the network prior to the final SoftMax layer
⚫Per sample, we compute the distance between V(sample) and V(train) for the training set
⚫Check how consistent the network is across the closest training examples
Labeled Dataset FC outputs
18 © NEC Corporation 2015 NEC Group Internal Use Only18 © NEC Corporation 2019
Abstention Criteria
Compare activations along
the feedforward path
19 © NEC Corporation 2015 NEC Group Internal Use Only19 © NEC Corporation 2019
Model Improvement
Find similar labeled cases
- Requires a fine ‘matching’ measure
- Works well only if labelled set is ‘super-dense’
- Hence – apply continuous deformable transformations from
the new sample to the closest training set cases
20 © NEC Corporation 2015 NEC Group Internal Use Only20 © NEC Corporation 2019
Explainablity and its relation to Augmentation
▌Given a set of legitimate transformations {Ei(X)}, where X is a labeled
input vector, an explanation for the classification of another vector Y is a
series of infinitesimal transformations which starts with the labeled
example X and ends in Y
▌The ‘shorter’ the series, the ‘better’ the explanation
▌Legitimate transformation: class of {Ei(X)}=class of X (or Y).
VS
IBM Research
21 © NEC Corporation 2015 NEC Group Internal Use Only21 © NEC Corporation 2019
Some Visualizations to the rescue
22 © NEC Corporation 2015 NEC Group Internal Use Only22 © NEC Corporation 2019
The path is not a trivial function..
▌For example, simple averaging is not OK
▌But applying some deformable image transformation is OK:
23 © NEC Corporation 2015 NEC Group Internal Use Only23 © NEC Corporation 2019
Luckily, Medical Imaging Research has the same challenge
▌Human organs change in time, and automatic registration is required
24 © NEC Corporation 2015 NEC Group Internal Use Only24 © NEC Corporation 2019
▌Using Lenet-5, for a new sample, we find a labeled candidate ‘nearest
neighbor’ using the before-last FC layer activations.
▌We apply a deformable DNN to ‘reach’ sample from training patch
▌One can estimate ‘reliability’ using number of activation flips along path
between labeled and new sample
Sandbox case – MNIST...
New Sample Labeled Sample
25 © NEC Corporation 2015 NEC Group Internal Use Only25 © NEC Corporation 2019
▌If new sample is mislabeled by the network, the ‘path’ is used as new
training samples.
▌Very fast convergence and higher accuracy reached easily.
Sandbox case – MNIST (cont.)
0 0 0 6 6 6 6 6 6 6
distance of FC layer from labelled image
(blue) and new image (red)
I. Aganj, J. E. Iglesias, M. Reuter, M. R. Sabuncu,
and B. Fischl, "Mid-space-independent deformable
image registration”
26 © NEC Corporation 2015 NEC Group Internal Use Only26 © NEC Corporation 2019
▌Train Model
▌Find ‘glaring error’ Y according to auditor/customer
▌Find labeled nearest neighbor(s) X in “last FC” space.
▌Create a legitimate transform X->Y
▌Label images along the transform
▌Re-train, repeat
Summary of the explainability=>model improvement cycle
7 7 7 7 7 7 X X 1 1
27 © NEC Corporation 2015 NEC Group Internal Use Only27 © NEC Corporation 2019
▌We seek to segment via match to similar training patch
▌We apply a deformable DNN to ‘reach’ sample
patch from training patch
▌If we can’t find a train+transform
path, the pixel is not on a blood vessel.
28X28 patch in
new sample
28X28 patch in
training set
Deformable
DNN
Sample Medical case - blood vessel segmentation
28 © NEC Corporation 2015 NEC Group Internal Use Only28 © NEC Corporation 2019
Blood vessel segmentation
29 © NEC Corporation 2015 NEC Group Internal Use Only29 © NEC Corporation 2019
Up Close, segmentation is not so easy
30 © NEC Corporation 2015 NEC Group Internal Use Only30 © NEC Corporation 2019
The ‘Human explanation’
▌The Ophthalmologist knows blood vessels are connected and continuous.
▌Sees the whole picture (pattern of vessel) and ‘completes’ by intuition
▌DNN sees a region, has no built-in concept of connectedness or larger scale.
31 © NEC Corporation 2015 NEC Group Internal Use Only31 © NEC Corporation 2019
Good augmentation can use a small training dataset
2D UNET Data
Each layer contains two 3 × 3 convolutions each
followed by a leaky rectified linear unit
(LeakyReLU), and then a 2 × 2 max pooling with
strides of two in each dimension. In the synthesis
path, each layer consists of an upconvolution of 2
× 2 by strides of two in each dimension, followed
by two 3 × 3 convolutions each followed by a
LeakyReLU. Shortcut connections from layers of
equal resolution in the analysis path provide the
essential high-resolution features to the synthesis
path. In the last layer a 1×1 convolution reduces
the number of output channels to the number of
label.
Train data – 2 images
Val data – 57 images
Test data – 20 images
Data Augmentation:
1. Elastic Distortions
2. Perspective Transforms
3. Size Preserving Rotations
4. Size Preserving Shearing
5. Cropping
6. Rotate
7. Flip
8. Zoom
9. Color
10. Contrast
Outcome Visualization
Network Result
32 © NEC Corporation 2015 NEC Group Internal Use Only32 © NEC Corporation 2019
Details of the solution (I)
▌Due to the combinatorial explosion, we search for the minimum size window
in which it seems a DNN can still reach SOTA results.
▌For 2D BV Segmentation it seems region of 64X64->24X24 has a plateau in
performance.
▌In addition to ‘default’ augmentations (rotation, scaling etc.) we apply
bicubic spline based deformable image registrations, e.g. Eppenhof, K.
A., & Pluim, J. P. (2018). Pulmonary ct registration through supervised learning with
convolutional neural networks. IEEE transactions on medical imaging.
33 © NEC Corporation 2015 NEC Group Internal Use Only33 © NEC Corporation 2019
Details of the solution (II)
▌For the new ‘test’ image, we find for each patch the most similar ‘existing
past case’. For that we use Locality-Similar Hashing, working on the actual
patch grayscale values as the input vector
▌With a 28X28=784 element vector, a hash size of 36 and 10 hash tables, we
do an LSH search with Euclidean distance measure.
▌For the first N=5 candidates, we compute the normalized cross correlation of
the training patch and the test patch and choose highest one.
▌An NCC >0.85 is considered a match.
▌If needed, for an even stronger match (NCC => 1.0) match, one can use
DNN based unsupervised deformable image registration.
de Vos, Bob D., et al. "End-to-end unsupervised deformable image registration with a
convolutional neural network." Deep Learning in Medical Image Analysis and Multimodal
Learning for Clinical Decision Support. Springer, Cham, 2017. 204-212.
34 © NEC Corporation 2015 NEC Group Internal Use Only34 © NEC Corporation 2019
Sanity checks
Train Sample – all white (explainable) Artificially Added digits – not explainable
35 © NEC Corporation 2015 NEC Group Internal Use Only35 © NEC Corporation 2019
Xplain-able AI in Cyber
▌Purpose: Automatic Anomaly Detection in SOC logs.
▌SOC engineers need to trust AI alerts
▌Solution: retrieve relevant training samples based on AutoEncoders, LIME to mark log
entries
36 © NEC Corporation 2015 NEC Group Internal Use Only36 © NEC Corporation 2019
Xplain-able AI in Edge AI
▌Purpose: Pedestrians/electric mobility analytics
▌Failure modes must be understood and predictable PRIOR to use
▌Solution: component-based detection + Synthetic Sample Generation
37 © NEC Corporation 2015 NEC Group Internal Use Only37 © NEC Corporation 2019
Xplain-able AI in Medical Diagnostics
38 © NEC Corporation 2015 NEC Group Internal Use Only38 © NEC Corporation 2019
AI in space
39 © NEC Corporation 2015 NEC Group Internal Use Only39 © NEC Corporation 2019
tsvi.lev@necam.com
Tsvi Lev. Practical Explainability for AI - with examples

More Related Content

Similar to Tsvi Lev. Practical Explainability for AI - with examples

ICLR 2020 Recap
ICLR 2020 RecapICLR 2020 Recap
ICLR 2020 Recap
Sri Ambati
 
PR-330: How To Train Your ViT? Data, Augmentation, and Regularization in Visi...
PR-330: How To Train Your ViT? Data, Augmentation, and Regularization in Visi...PR-330: How To Train Your ViT? Data, Augmentation, and Regularization in Visi...
PR-330: How To Train Your ViT? Data, Augmentation, and Regularization in Visi...
Jinwon Lee
 

Similar to Tsvi Lev. Practical Explainability for AI - with examples (20)

CAR DAMAGE DETECTION USING DEEP LEARNING
CAR DAMAGE DETECTION USING DEEP LEARNINGCAR DAMAGE DETECTION USING DEEP LEARNING
CAR DAMAGE DETECTION USING DEEP LEARNING
 
"Deep Learning Beyond Cats and Cars: Developing a Real-life DNN-based Embedde...
"Deep Learning Beyond Cats and Cars: Developing a Real-life DNN-based Embedde..."Deep Learning Beyond Cats and Cars: Developing a Real-life DNN-based Embedde...
"Deep Learning Beyond Cats and Cars: Developing a Real-life DNN-based Embedde...
 
深度學習在AOI的應用
深度學習在AOI的應用深度學習在AOI的應用
深度學習在AOI的應用
 
IRJET- A Study of Different Convolution Neural Network Architectures for Huma...
IRJET- A Study of Different Convolution Neural Network Architectures for Huma...IRJET- A Study of Different Convolution Neural Network Architectures for Huma...
IRJET- A Study of Different Convolution Neural Network Architectures for Huma...
 
ICLR 2020 Recap
ICLR 2020 RecapICLR 2020 Recap
ICLR 2020 Recap
 
ams’ NanEye Mini Camera
ams’ NanEye Mini Cameraams’ NanEye Mini Camera
ams’ NanEye Mini Camera
 
AICamp - Dr Ramine Tinati - Making Computer Vision Real
AICamp - Dr Ramine Tinati - Making Computer Vision RealAICamp - Dr Ramine Tinati - Making Computer Vision Real
AICamp - Dr Ramine Tinati - Making Computer Vision Real
 
Build computer vision models to perform object detection and classification w...
Build computer vision models to perform object detection and classification w...Build computer vision models to perform object detection and classification w...
Build computer vision models to perform object detection and classification w...
 
Programming with Relaxed Synchronization
Programming with Relaxed SynchronizationProgramming with Relaxed Synchronization
Programming with Relaxed Synchronization
 
“An Industry Standard Performance Benchmark Suite for Machine Learning,” a Pr...
“An Industry Standard Performance Benchmark Suite for Machine Learning,” a Pr...“An Industry Standard Performance Benchmark Suite for Machine Learning,” a Pr...
“An Industry Standard Performance Benchmark Suite for Machine Learning,” a Pr...
 
Traffic Sign Recognition System
Traffic Sign Recognition SystemTraffic Sign Recognition System
Traffic Sign Recognition System
 
Comparative Study of Pre-Trained Neural Network Models in Detection of Glaucoma
Comparative Study of Pre-Trained Neural Network Models in Detection of GlaucomaComparative Study of Pre-Trained Neural Network Models in Detection of Glaucoma
Comparative Study of Pre-Trained Neural Network Models in Detection of Glaucoma
 
The Overture ΔQ testbed for design and deployment planning
The Overture ΔQ testbed for design and deployment planningThe Overture ΔQ testbed for design and deployment planning
The Overture ΔQ testbed for design and deployment planning
 
IRJET - Autonomous Navigation System using Deep Learning
IRJET -  	  Autonomous Navigation System using Deep LearningIRJET -  	  Autonomous Navigation System using Deep Learning
IRJET - Autonomous Navigation System using Deep Learning
 
Introduction to Machine Learning on IBM Power Systems
Introduction to Machine Learning on IBM Power SystemsIntroduction to Machine Learning on IBM Power Systems
Introduction to Machine Learning on IBM Power Systems
 
PR-330: How To Train Your ViT? Data, Augmentation, and Regularization in Visi...
PR-330: How To Train Your ViT? Data, Augmentation, and Regularization in Visi...PR-330: How To Train Your ViT? Data, Augmentation, and Regularization in Visi...
PR-330: How To Train Your ViT? Data, Augmentation, and Regularization in Visi...
 
IRJET- Spatial Context Preservation and Propagation - Layer States in Convolu...
IRJET- Spatial Context Preservation and Propagation - Layer States in Convolu...IRJET- Spatial Context Preservation and Propagation - Layer States in Convolu...
IRJET- Spatial Context Preservation and Propagation - Layer States in Convolu...
 
Operationalizing SDN
Operationalizing SDNOperationalizing SDN
Operationalizing SDN
 
SDNOrigins.ppt
SDNOrigins.pptSDNOrigins.ppt
SDNOrigins.ppt
 
Practical Artificial Intelligence: Deep Learning Beyond Cats and Cars
Practical Artificial Intelligence: Deep Learning Beyond Cats and CarsPractical Artificial Intelligence: Deep Learning Beyond Cats and Cars
Practical Artificial Intelligence: Deep Learning Beyond Cats and Cars
 

More from Lviv Startup Club

More from Lviv Startup Club (20)

Artem Bykovets: 4 Вершники апокаліпсису робочих стосунків (+антидоти до них) ...
Artem Bykovets: 4 Вершники апокаліпсису робочих стосунків (+антидоти до них) ...Artem Bykovets: 4 Вершники апокаліпсису робочих стосунків (+антидоти до них) ...
Artem Bykovets: 4 Вершники апокаліпсису робочих стосунків (+антидоти до них) ...
 
Dmytro Khudenko: Challenges of implementing task managers in the corporate an...
Dmytro Khudenko: Challenges of implementing task managers in the corporate an...Dmytro Khudenko: Challenges of implementing task managers in the corporate an...
Dmytro Khudenko: Challenges of implementing task managers in the corporate an...
 
Sergii Melnichenko: Лідерство в Agile командах: ТОП-5 основних психологічних ...
Sergii Melnichenko: Лідерство в Agile командах: ТОП-5 основних психологічних ...Sergii Melnichenko: Лідерство в Agile командах: ТОП-5 основних психологічних ...
Sergii Melnichenko: Лідерство в Agile командах: ТОП-5 основних психологічних ...
 
Mariia Rashkevych: Підвищення ефективності розроблення та реалізації освітніх...
Mariia Rashkevych: Підвищення ефективності розроблення та реалізації освітніх...Mariia Rashkevych: Підвищення ефективності розроблення та реалізації освітніх...
Mariia Rashkevych: Підвищення ефективності розроблення та реалізації освітніх...
 
Mykhailo Hryhorash: What can be good in a "bad" project? (UA)
Mykhailo Hryhorash: What can be good in a "bad" project? (UA)Mykhailo Hryhorash: What can be good in a "bad" project? (UA)
Mykhailo Hryhorash: What can be good in a "bad" project? (UA)
 
Oleksii Kyselov: Що заважає ПМу зростати? Розбір практичних кейсів (UA)
Oleksii Kyselov: Що заважає ПМу зростати? Розбір практичних кейсів (UA)Oleksii Kyselov: Що заважає ПМу зростати? Розбір практичних кейсів (UA)
Oleksii Kyselov: Що заважає ПМу зростати? Розбір практичних кейсів (UA)
 
Yaroslav Osolikhin: «Неідеальний» проєктний менеджер: People Management під ч...
Yaroslav Osolikhin: «Неідеальний» проєктний менеджер: People Management під ч...Yaroslav Osolikhin: «Неідеальний» проєктний менеджер: People Management під ч...
Yaroslav Osolikhin: «Неідеальний» проєктний менеджер: People Management під ч...
 
Mariya Yeremenko: Вплив Генеративного ШІ на сучасний світ та на особисту ефек...
Mariya Yeremenko: Вплив Генеративного ШІ на сучасний світ та на особисту ефек...Mariya Yeremenko: Вплив Генеративного ШІ на сучасний світ та на особисту ефек...
Mariya Yeremenko: Вплив Генеративного ШІ на сучасний світ та на особисту ефек...
 
Petro Nikolaiev & Dmytro Kisov: ТОП-5 методів дослідження клієнтів для успіху...
Petro Nikolaiev & Dmytro Kisov: ТОП-5 методів дослідження клієнтів для успіху...Petro Nikolaiev & Dmytro Kisov: ТОП-5 методів дослідження клієнтів для успіху...
Petro Nikolaiev & Dmytro Kisov: ТОП-5 методів дослідження клієнтів для успіху...
 
Maksym Stelmakh : Державні електронні послуги та сервіси: чому бізнесу варто ...
Maksym Stelmakh : Державні електронні послуги та сервіси: чому бізнесу варто ...Maksym Stelmakh : Державні електронні послуги та сервіси: чому бізнесу варто ...
Maksym Stelmakh : Державні електронні послуги та сервіси: чому бізнесу варто ...
 
Alexander Marchenko: Проблеми росту продуктової екосистеми (UA)
Alexander Marchenko: Проблеми росту продуктової екосистеми (UA)Alexander Marchenko: Проблеми росту продуктової екосистеми (UA)
Alexander Marchenko: Проблеми росту продуктової екосистеми (UA)
 
Oleksandr Grytsenko: Save your Job або прокачай скіли до Engineering Manageme...
Oleksandr Grytsenko: Save your Job або прокачай скіли до Engineering Manageme...Oleksandr Grytsenko: Save your Job або прокачай скіли до Engineering Manageme...
Oleksandr Grytsenko: Save your Job або прокачай скіли до Engineering Manageme...
 
Yuliia Pieskova: Фідбек: не лише "як", але й "коли" і "навіщо" (UA)
Yuliia Pieskova: Фідбек: не лише "як", але й "коли" і "навіщо" (UA)Yuliia Pieskova: Фідбек: не лише "як", але й "коли" і "навіщо" (UA)
Yuliia Pieskova: Фідбек: не лише "як", але й "коли" і "навіщо" (UA)
 
Nataliya Kryvonis: Essential soft skills to lead your team (UA)
Nataliya Kryvonis: Essential soft skills to lead your team (UA)Nataliya Kryvonis: Essential soft skills to lead your team (UA)
Nataliya Kryvonis: Essential soft skills to lead your team (UA)
 
Volodymyr Salyha: Stakeholder Alchemy: Transforming Analysis into Meaningful ...
Volodymyr Salyha: Stakeholder Alchemy: Transforming Analysis into Meaningful ...Volodymyr Salyha: Stakeholder Alchemy: Transforming Analysis into Meaningful ...
Volodymyr Salyha: Stakeholder Alchemy: Transforming Analysis into Meaningful ...
 
Anna Chalyuk: 7 інструментів та принципів, які допоможуть зробити вашу команд...
Anna Chalyuk: 7 інструментів та принципів, які допоможуть зробити вашу команд...Anna Chalyuk: 7 інструментів та принципів, які допоможуть зробити вашу команд...
Anna Chalyuk: 7 інструментів та принципів, які допоможуть зробити вашу команд...
 
Oksana Smilka: Цінності, цілі та (де) мотивація (UA)
Oksana Smilka: Цінності, цілі та (де) мотивація (UA)Oksana Smilka: Цінності, цілі та (де) мотивація (UA)
Oksana Smilka: Цінності, цілі та (де) мотивація (UA)
 
Yaroslav Rozhankivskyy: Три складові і три передумови максимальної продуктивн...
Yaroslav Rozhankivskyy: Три складові і три передумови максимальної продуктивн...Yaroslav Rozhankivskyy: Три складові і три передумови максимальної продуктивн...
Yaroslav Rozhankivskyy: Три складові і три передумови максимальної продуктивн...
 
Andrii Skoromnyi: Чому не працює методика "5 Чому?" – і яка є альтернатива? (UA)
Andrii Skoromnyi: Чому не працює методика "5 Чому?" – і яка є альтернатива? (UA)Andrii Skoromnyi: Чому не працює методика "5 Чому?" – і яка є альтернатива? (UA)
Andrii Skoromnyi: Чому не працює методика "5 Чому?" – і яка є альтернатива? (UA)
 
Maryna Sokyrko & Oleksandr Chugui: Building Product Passion: Developing AI ch...
Maryna Sokyrko & Oleksandr Chugui: Building Product Passion: Developing AI ch...Maryna Sokyrko & Oleksandr Chugui: Building Product Passion: Developing AI ch...
Maryna Sokyrko & Oleksandr Chugui: Building Product Passion: Developing AI ch...
 

Recently uploaded

會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文
會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文
會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文
中 央社
 
會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽
會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽
會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽
中 央社
 
Transparency, Recognition and the role of eSealing - Ildiko Mazar and Koen No...
Transparency, Recognition and the role of eSealing - Ildiko Mazar and Koen No...Transparency, Recognition and the role of eSealing - Ildiko Mazar and Koen No...
Transparency, Recognition and the role of eSealing - Ildiko Mazar and Koen No...
EADTU
 
Spellings Wk 4 and Wk 5 for Grade 4 at CAPS
Spellings Wk 4 and Wk 5 for Grade 4 at CAPSSpellings Wk 4 and Wk 5 for Grade 4 at CAPS
Spellings Wk 4 and Wk 5 for Grade 4 at CAPS
AnaAcapella
 

Recently uploaded (20)

Mattingly "AI and Prompt Design: LLMs with NER"
Mattingly "AI and Prompt Design: LLMs with NER"Mattingly "AI and Prompt Design: LLMs with NER"
Mattingly "AI and Prompt Design: LLMs with NER"
 
會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文
會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文
會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文會考英文
 
會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽
會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽
會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽會考英聽
 
ĐỀ THAM KHẢO KÌ THI TUYỂN SINH VÀO LỚP 10 MÔN TIẾNG ANH FORM 50 CÂU TRẮC NGHI...
ĐỀ THAM KHẢO KÌ THI TUYỂN SINH VÀO LỚP 10 MÔN TIẾNG ANH FORM 50 CÂU TRẮC NGHI...ĐỀ THAM KHẢO KÌ THI TUYỂN SINH VÀO LỚP 10 MÔN TIẾNG ANH FORM 50 CÂU TRẮC NGHI...
ĐỀ THAM KHẢO KÌ THI TUYỂN SINH VÀO LỚP 10 MÔN TIẾNG ANH FORM 50 CÂU TRẮC NGHI...
 
Mattingly "AI & Prompt Design: Named Entity Recognition"
Mattingly "AI & Prompt Design: Named Entity Recognition"Mattingly "AI & Prompt Design: Named Entity Recognition"
Mattingly "AI & Prompt Design: Named Entity Recognition"
 
8 Tips for Effective Working Capital Management
8 Tips for Effective Working Capital Management8 Tips for Effective Working Capital Management
8 Tips for Effective Working Capital Management
 
VAMOS CUIDAR DO NOSSO PLANETA! .
VAMOS CUIDAR DO NOSSO PLANETA!                    .VAMOS CUIDAR DO NOSSO PLANETA!                    .
VAMOS CUIDAR DO NOSSO PLANETA! .
 
AIM of Education-Teachers Training-2024.ppt
AIM of Education-Teachers Training-2024.pptAIM of Education-Teachers Training-2024.ppt
AIM of Education-Teachers Training-2024.ppt
 
OSCM Unit 2_Operations Processes & Systems
OSCM Unit 2_Operations Processes & SystemsOSCM Unit 2_Operations Processes & Systems
OSCM Unit 2_Operations Processes & Systems
 
male presentation...pdf.................
male presentation...pdf.................male presentation...pdf.................
male presentation...pdf.................
 
The Story of Village Palampur Class 9 Free Study Material PDF
The Story of Village Palampur Class 9 Free Study Material PDFThe Story of Village Palampur Class 9 Free Study Material PDF
The Story of Village Palampur Class 9 Free Study Material PDF
 
Transparency, Recognition and the role of eSealing - Ildiko Mazar and Koen No...
Transparency, Recognition and the role of eSealing - Ildiko Mazar and Koen No...Transparency, Recognition and the role of eSealing - Ildiko Mazar and Koen No...
Transparency, Recognition and the role of eSealing - Ildiko Mazar and Koen No...
 
e-Sealing at EADTU by Kamakshi Rajagopal
e-Sealing at EADTU by Kamakshi Rajagopale-Sealing at EADTU by Kamakshi Rajagopal
e-Sealing at EADTU by Kamakshi Rajagopal
 
When Quality Assurance Meets Innovation in Higher Education - Report launch w...
When Quality Assurance Meets Innovation in Higher Education - Report launch w...When Quality Assurance Meets Innovation in Higher Education - Report launch w...
When Quality Assurance Meets Innovation in Higher Education - Report launch w...
 
Spellings Wk 4 and Wk 5 for Grade 4 at CAPS
Spellings Wk 4 and Wk 5 for Grade 4 at CAPSSpellings Wk 4 and Wk 5 for Grade 4 at CAPS
Spellings Wk 4 and Wk 5 for Grade 4 at CAPS
 
Basic Civil Engineering notes on Transportation Engineering & Modes of Transport
Basic Civil Engineering notes on Transportation Engineering & Modes of TransportBasic Civil Engineering notes on Transportation Engineering & Modes of Transport
Basic Civil Engineering notes on Transportation Engineering & Modes of Transport
 
FICTIONAL SALESMAN/SALESMAN SNSW 2024.pdf
FICTIONAL SALESMAN/SALESMAN SNSW 2024.pdfFICTIONAL SALESMAN/SALESMAN SNSW 2024.pdf
FICTIONAL SALESMAN/SALESMAN SNSW 2024.pdf
 
OS-operating systems- ch05 (CPU Scheduling) ...
OS-operating systems- ch05 (CPU Scheduling) ...OS-operating systems- ch05 (CPU Scheduling) ...
OS-operating systems- ch05 (CPU Scheduling) ...
 
ANTI PARKISON DRUGS.pptx
ANTI         PARKISON          DRUGS.pptxANTI         PARKISON          DRUGS.pptx
ANTI PARKISON DRUGS.pptx
 
Stl Algorithms in C++ jjjjjjjjjjjjjjjjjj
Stl Algorithms in C++ jjjjjjjjjjjjjjjjjjStl Algorithms in C++ jjjjjjjjjjjjjjjjjj
Stl Algorithms in C++ jjjjjjjjjjjjjjjjjj
 

Tsvi Lev. Practical Explainability for AI - with examples

  • 1. Tsvi Lev, Corporate VP and Head of NEC Labs Israel tsvi.lev@necam.com XAI Beyond the Hype #MAXA
  • 2. 2 © NEC Corporation 2015 NEC Group Internal Use Only2 © NEC Corporation 2019 NEC – we too were once a startup..
  • 3. 3 © NEC Corporation 2015 NEC Group Internal Use Only3 © NEC Corporation 2019 120 Years of Research Innovation
  • 4. 4 © NEC Corporation 2015 NEC Group Internal Use Only4 © NEC Corporation 2019 Applied AI is quite different from Kaggle competitions or papers NEC Labs do Applied AI for vision, healthcare, cyber, edge. Our Customers: ▌don’t care about SOTA or AUC ▌lack training data ▌want AI to behave ‘like us’ ACM CCS Best Paper Edge AI for Vision Cyber Research AI for Healthcare
  • 5. 5 © NEC Corporation 2015 NEC Group Internal Use Only5 © NEC Corporation 2019 ▌ The recent AI craze started with Digital Marketing success ($$$) and Academic breakthroughs in Image and Speech Processing ▌Other application areas have different requirements The truth about recent AI success More $$$! Better than Human
  • 6. 6 © NEC Corporation 2015 NEC Group Internal Use Only6 © NEC Corporation 2019 ▌ Recommendation Engines replaced rule based systems ▌ Measured only by results (clicks, sales) ▌ Recently – issues of Bias and Accountability Use Case: digital marketing Accuracy Stability Compliance Explainablity Accountability Bias Free
  • 7. 7 © NEC Corporation 2015 NEC Group Internal Use Only7 © NEC Corporation 2019 ▌Replace non-expert Humans ▌Acceptance threshold prevents ‘just better than previous’ criteria ▌Stability and Compliance are key from day 0 ▌Hence – slower than promised and hyped Use Case: Mobility Accuracy Stability Compliance Explainablity Accountability Bias Free
  • 8. 8 © NEC Corporation 2015 NEC Group Internal Use Only8 © NEC Corporation 2019 ▌ Assist, not replace Expert Human Judgement. ▌ Xplain-ability and Stability are >> than accuracy ▌ Hence – slower, but progress is stable Use Case : Medical Diagnostics Accuracy Stability Compliance Explainablity Accountability Bias Free
  • 9. 9 © NEC Corporation 2015 NEC Group Internal Use Only9 © NEC Corporation 2019 The Future will be dominated by XAI “Confidence in AI only comes with explainablity and transparency” – Well Known B2B AI vendor
  • 10. 10 © NEC Corporation 2015 NEC Group Internal Use Only10 © NEC Corporation 2019 When do we most need XAI the most? Recommendation Engines Man Machine Interfaces Online Advertising Gaming Single Decision=> Lives, much $$$ Many Micro Decision Aggregated
  • 11. 11 © NEC Corporation 2015 NEC Group Internal Use Only11 © NEC Corporation 2019 What may happen without XAI – AUC/ROC deterioration Populations, physicians, imaging machines change – how to ensure AI is constant? Benign Malignant! Benign
  • 12. 12 © NEC Corporation 2015 NEC Group Internal Use Only12 © NEC Corporation 2019 Users attach different gravity to certain error cases ▌ Errors in “simple” cases: blood vessel vs nodules ▌ Distribution of errors: no false alarms, then suddenly a ton for no reason ▌ “Butterfly in Asia” – tiny changes in input yield large changes in the result
  • 13. 13 © NEC Corporation 2015 NEC Group Internal Use Only13 © NEC Corporation 2019 Which Models are a best fit for XAI?
  • 14. 14 © NEC Corporation 2015 NEC Group Internal Use Only14 © NEC Corporation 2019 Model type does not guarantee explainability MNIST decision tree (Hinton et al)
  • 15. 15 © NEC Corporation 2015 NEC Group Internal Use Only15 © NEC Corporation 2019 Explainable can mean different things Causality Analysis Reliability/Abstention Model Improvement Pattern Discovery Post-hoc: methods to be applied to other models. IBM Research
  • 16. 16 © NEC Corporation 2015 NEC Group Internal Use Only16 © NEC Corporation 2019 Causality Analysis Local Interpretable Model-agnostic Explanations (LIME)
  • 17. 17 © NEC Corporation 2015 NEC Group Internal Use Only17 © NEC Corporation 2019 Reliability/Abstention Criteria ▌Threshold – estimate the reliability by the output magnitude ▌Distance to labeled examples – ⚫Let V be the output vector of the network prior to the final SoftMax layer ⚫Per sample, we compute the distance between V(sample) and V(train) for the training set ⚫Check how consistent the network is across the closest training examples Labeled Dataset FC outputs
  • 18. 18 © NEC Corporation 2015 NEC Group Internal Use Only18 © NEC Corporation 2019 Abstention Criteria Compare activations along the feedforward path
  • 19. 19 © NEC Corporation 2015 NEC Group Internal Use Only19 © NEC Corporation 2019 Model Improvement Find similar labeled cases - Requires a fine ‘matching’ measure - Works well only if labelled set is ‘super-dense’ - Hence – apply continuous deformable transformations from the new sample to the closest training set cases
  • 20. 20 © NEC Corporation 2015 NEC Group Internal Use Only20 © NEC Corporation 2019 Explainablity and its relation to Augmentation ▌Given a set of legitimate transformations {Ei(X)}, where X is a labeled input vector, an explanation for the classification of another vector Y is a series of infinitesimal transformations which starts with the labeled example X and ends in Y ▌The ‘shorter’ the series, the ‘better’ the explanation ▌Legitimate transformation: class of {Ei(X)}=class of X (or Y). VS IBM Research
  • 21. 21 © NEC Corporation 2015 NEC Group Internal Use Only21 © NEC Corporation 2019 Some Visualizations to the rescue
  • 22. 22 © NEC Corporation 2015 NEC Group Internal Use Only22 © NEC Corporation 2019 The path is not a trivial function.. ▌For example, simple averaging is not OK ▌But applying some deformable image transformation is OK:
  • 23. 23 © NEC Corporation 2015 NEC Group Internal Use Only23 © NEC Corporation 2019 Luckily, Medical Imaging Research has the same challenge ▌Human organs change in time, and automatic registration is required
  • 24. 24 © NEC Corporation 2015 NEC Group Internal Use Only24 © NEC Corporation 2019 ▌Using Lenet-5, for a new sample, we find a labeled candidate ‘nearest neighbor’ using the before-last FC layer activations. ▌We apply a deformable DNN to ‘reach’ sample from training patch ▌One can estimate ‘reliability’ using number of activation flips along path between labeled and new sample Sandbox case – MNIST... New Sample Labeled Sample
  • 25. 25 © NEC Corporation 2015 NEC Group Internal Use Only25 © NEC Corporation 2019 ▌If new sample is mislabeled by the network, the ‘path’ is used as new training samples. ▌Very fast convergence and higher accuracy reached easily. Sandbox case – MNIST (cont.) 0 0 0 6 6 6 6 6 6 6 distance of FC layer from labelled image (blue) and new image (red) I. Aganj, J. E. Iglesias, M. Reuter, M. R. Sabuncu, and B. Fischl, "Mid-space-independent deformable image registration”
  • 26. 26 © NEC Corporation 2015 NEC Group Internal Use Only26 © NEC Corporation 2019 ▌Train Model ▌Find ‘glaring error’ Y according to auditor/customer ▌Find labeled nearest neighbor(s) X in “last FC” space. ▌Create a legitimate transform X->Y ▌Label images along the transform ▌Re-train, repeat Summary of the explainability=>model improvement cycle 7 7 7 7 7 7 X X 1 1
  • 27. 27 © NEC Corporation 2015 NEC Group Internal Use Only27 © NEC Corporation 2019 ▌We seek to segment via match to similar training patch ▌We apply a deformable DNN to ‘reach’ sample patch from training patch ▌If we can’t find a train+transform path, the pixel is not on a blood vessel. 28X28 patch in new sample 28X28 patch in training set Deformable DNN Sample Medical case - blood vessel segmentation
  • 28. 28 © NEC Corporation 2015 NEC Group Internal Use Only28 © NEC Corporation 2019 Blood vessel segmentation
  • 29. 29 © NEC Corporation 2015 NEC Group Internal Use Only29 © NEC Corporation 2019 Up Close, segmentation is not so easy
  • 30. 30 © NEC Corporation 2015 NEC Group Internal Use Only30 © NEC Corporation 2019 The ‘Human explanation’ ▌The Ophthalmologist knows blood vessels are connected and continuous. ▌Sees the whole picture (pattern of vessel) and ‘completes’ by intuition ▌DNN sees a region, has no built-in concept of connectedness or larger scale.
  • 31. 31 © NEC Corporation 2015 NEC Group Internal Use Only31 © NEC Corporation 2019 Good augmentation can use a small training dataset 2D UNET Data Each layer contains two 3 × 3 convolutions each followed by a leaky rectified linear unit (LeakyReLU), and then a 2 × 2 max pooling with strides of two in each dimension. In the synthesis path, each layer consists of an upconvolution of 2 × 2 by strides of two in each dimension, followed by two 3 × 3 convolutions each followed by a LeakyReLU. Shortcut connections from layers of equal resolution in the analysis path provide the essential high-resolution features to the synthesis path. In the last layer a 1×1 convolution reduces the number of output channels to the number of label. Train data – 2 images Val data – 57 images Test data – 20 images Data Augmentation: 1. Elastic Distortions 2. Perspective Transforms 3. Size Preserving Rotations 4. Size Preserving Shearing 5. Cropping 6. Rotate 7. Flip 8. Zoom 9. Color 10. Contrast Outcome Visualization Network Result
  • 32. 32 © NEC Corporation 2015 NEC Group Internal Use Only32 © NEC Corporation 2019 Details of the solution (I) ▌Due to the combinatorial explosion, we search for the minimum size window in which it seems a DNN can still reach SOTA results. ▌For 2D BV Segmentation it seems region of 64X64->24X24 has a plateau in performance. ▌In addition to ‘default’ augmentations (rotation, scaling etc.) we apply bicubic spline based deformable image registrations, e.g. Eppenhof, K. A., & Pluim, J. P. (2018). Pulmonary ct registration through supervised learning with convolutional neural networks. IEEE transactions on medical imaging.
  • 33. 33 © NEC Corporation 2015 NEC Group Internal Use Only33 © NEC Corporation 2019 Details of the solution (II) ▌For the new ‘test’ image, we find for each patch the most similar ‘existing past case’. For that we use Locality-Similar Hashing, working on the actual patch grayscale values as the input vector ▌With a 28X28=784 element vector, a hash size of 36 and 10 hash tables, we do an LSH search with Euclidean distance measure. ▌For the first N=5 candidates, we compute the normalized cross correlation of the training patch and the test patch and choose highest one. ▌An NCC >0.85 is considered a match. ▌If needed, for an even stronger match (NCC => 1.0) match, one can use DNN based unsupervised deformable image registration. de Vos, Bob D., et al. "End-to-end unsupervised deformable image registration with a convolutional neural network." Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support. Springer, Cham, 2017. 204-212.
  • 34. 34 © NEC Corporation 2015 NEC Group Internal Use Only34 © NEC Corporation 2019 Sanity checks Train Sample – all white (explainable) Artificially Added digits – not explainable
  • 35. 35 © NEC Corporation 2015 NEC Group Internal Use Only35 © NEC Corporation 2019 Xplain-able AI in Cyber ▌Purpose: Automatic Anomaly Detection in SOC logs. ▌SOC engineers need to trust AI alerts ▌Solution: retrieve relevant training samples based on AutoEncoders, LIME to mark log entries
  • 36. 36 © NEC Corporation 2015 NEC Group Internal Use Only36 © NEC Corporation 2019 Xplain-able AI in Edge AI ▌Purpose: Pedestrians/electric mobility analytics ▌Failure modes must be understood and predictable PRIOR to use ▌Solution: component-based detection + Synthetic Sample Generation
  • 37. 37 © NEC Corporation 2015 NEC Group Internal Use Only37 © NEC Corporation 2019 Xplain-able AI in Medical Diagnostics
  • 38. 38 © NEC Corporation 2015 NEC Group Internal Use Only38 © NEC Corporation 2019 AI in space
  • 39. 39 © NEC Corporation 2015 NEC Group Internal Use Only39 © NEC Corporation 2019 tsvi.lev@necam.com