ILLUMINATING AI: UNDERSTANDING AI'S GOALS, REASONING & - - PowerPoint PPT Presentation

ILLUMINATING AI:

UNDERSTANDING AI'S GOALS, REASONING & COMPROMISES

AI says “7”

(99%)

Tsvi Achler MD PhD Tsvi Achler MD PhD Image

If SIRI makes a mistake, the impact is limited

AI Adoption Requires Transparency: Trust, Regulation, and Understanding of the AI’s Compromises But the problem is: AI is a Black Box

2

Banking | Medicine | Self-Driving Cars

In most applications a mistake has more serious consequences:

With undetectable noise Google

+ =

The Lack of Transparency Leaves You Asking: What Is the AI Really Recognizing? DARPA commitment: “Explainable AI” Initiative

3

EU Legislates: Users Have a Right to an Explanation

Before

Solution Pathways for Explanability (1) DARPA:

Trial & Error to Find What Effects the Network Mirrors the brain’s ability to provide reasons

(2) Optimizing Mind:

Takes Time & the AI Remains a Black Box The Ground Truth Of The AI

5

The Brain Relies on Feedback

During Recognition

AI Does Not

Feedback: No Feedback:

Feedback is Found Throughout the Brain

(eg Aroniadou-Anderjaska et al 2000)

Tri-synaptic connections

More Feedback Than Feedforward

Retrograde Signaling e.g.: nitric oxide

AI Exclusively Uses Feedforward Connections W During Recognition

Outputs Y Inputs X

Feedforward Caricature Computational Caricature Mathematical Function

W=

Outputs Y Inputs X

Connectivity Notation Weight Matrix

Y X

W

X W Y   

(during recognition)

Recognition

8

Do not be mislead:

Even when an AI is called

“recurrent” it still uses W

Why Is Feedback Needed? For Optimization

• What is Optimization?
• Difference between “Feedforward” and

Feedforward-Feedback methods

• Why lack of Feedback During Recognition matters

9

Optimization

Try configuration, evaluate, modify and repeat until optimal fit

Example: Solving Jigsaw Puzzle

(OP)

OP

Try Evaluate Modify

10

OP

Recognition Algorithms:

Y X

Bird Bird

Bird

Recall Recognition- Inference Learning Memory Learned feedforward weights in: Deep, Convolutional, Recurrent, LSTM, Reinforcement Networks, Support Vector Machines (SVM), Perceptrons, “Neural Networks” … everything learned via Backprop etc.

by Optimizing (OP) during learning

W

Fodor & Pylyshyn (1988) Sun (2002)

WX Y 

OP

Optimize weights so that recognition occurs using a simple multiplication 1) Optimized weights W are a “Black Box “Feedforward” methods

OP

Recognition Algorithms:

Y X

Bird Bird

Bird

Recall Recognition- Inference Learning Memory

W

Fodor & Pylyshyn (1988) Sun (2002)

WX Y 

OP

encodes uniqueness into weights “Feedforward” methods

Determining Uniqueness is essential to perform efficient recognition Problem: Uniqueness changes with context cannot learn uniqueness for all possible contexts

O O O O O X X X O X O

Besides: relevant context is during recognition, not learning

Unique thus Important! Unique thus Important! Training Instance 1 Training Instance 2

OP

We suggest uniqueness is determined during recognition instead

Bird Bird

Bird

Recall Recognition Learning Memory

2) Optimizing only current test pattern Weights “Clear Box” 1) Determining activation Y (not weights) → Reducing computational costs OP (not all of training data) By Optimizing (OP) during recognition

when the context is available

while estimating uniqueness

15

Model-type During Learning (weight Δ) During Recognition (find Y)

Why would the brain only use feedback during learning?

Optimization “Feedforward” Feedforward recognition Simpler Learning M Illuminated AI to find weights W Switch Dynamics to find neuron activation Optimization

Illuminated AI Switches When Optimization Occurs

Requires feedback for learning: for example to back-propagate error Not really Feedforward !!!

Recognition with Illuminated AI

Outputs Y Inputs X

Symmetrical Inhibitory connections modulate inputs using output activity

M=

Outputs Y Inputs X

Y X

M

M

For optimization Weights are Expectations (allows explainability & update)

Neuron Caricature Computational Caricature Connectivity Notation

17

Same Results But Transparent Method

X W Y   

pattern from the environment (input) resulting neuron activation (output) "feedforward" weights

SAME SAME

M

Illuminated Networks:

Y  X 

OP

Feedforward:

SAME

Y  X  Y 

18

Easier to Explain, Learn and Update Illuminated Weights

Example: You train your AI It gets good grades (performance) You are done … right?

Learn Digits

95%

19

1) Can convert existing feedforward networks to xRFN & see what they are doing: Black Box -> “Clear Box”

MNIST Demonstration

SVM Feedforward xRFN

See Inside!

Equivalent

Overall Accuracy 91.65% By Digit: 1 2 3 4 5 6 7 8 9 False Positives 45 67 106 79 86 67 75 152 109 49 False Negatives 19 129 91 71 137 50 83 125 109 21 SVM Overall Accuracy 91.65% By Digit: 1 2 3 4 5 6 7 8 9 False Positives 45 67 106 79 86 67 75 152 109 49 False Negatives 19 129 91 71 137 50 83 125 109 21 RFN

Why are Explainable Regulatory Feedback Networks xRFN beneficial?

Does the brain perform optimization during recognition? O X O O O O O O O O O O O O O O O O O O O O O E E E E E E E E E E E E E E E E E E E F vs.

Rosenholtz 2001 This occurs in all modalities, including audition, vision, tactile, and (like seen in the jigsaw puzzle)

suggests optimization during recognition is ubiquitous

Rinberg etal 2006

How long does it take to Find the single pattern?

even in olfaction with its poor spatial resolution If brain uses optimization: should be faster with unique patterns (like jigsaw) If brain uses Feedforwad: Y=WX fixed propagation, fixed amount of time (taking what is commonly considered a spatial attention phenomena and attributing it to a recognition phenomena) Brain takes longer in right box suggesting a signal-to-noise phenomena

Does It Scale To Large AI?

Does Illuminated AI Consume More Resources Than Feedforward AI?

Tests on Random Data Increasing in Size

Nodes Features Matrix size 10 100 1,000 100 1,000 100,000 500 1,000 500,000 1,000 10,000 10,000,000 2,000 10,000 20,000,000 6,000 12,000 72,000,000 8,000 15,000 120,000,000 9,000 20,000 180,000,000

1,000 2,000 3,000 4,000 5,000 6,000 7,000 Computing Time (s)

Computational Costs:

During Learning

SVM Learning (W) Fastest Feedforward Learning (W) Illuminated Learning (M)

Out of Memory!

5 10 15 20 Matrix Size Million

Out of Memory!

… 120 !

SVM F

Can Learn > 100x Faster Without Balancing Data

Out of Memory!

2 4 6 8 10 Computational Cost per Test (s)

Computational Costs:

During Recognition

Best Alternate AI (KNN)

Without Optimization

Illuminated AI (M) Matrix Size 50 100 Millions 20 120 Feedforward AI (W)

Nodes Features Matrix size 10 100 1,000 100 1,000 100,000 500 1,000 500,000 1,000 10,000 10,000,000 2,000 10,000 20,000,000 6,000 12,000 72,000,000 8,000 15,000 120,000,000 9,000 20,000 180,000,000 SVM F Out of Memory!

Accelerated on GPU’s

25

Torch/Lua

Also Useful for Simpler AI Such As: Logistic Regression & Random Forests

26

The Current Standard For Explainability Is Decision Trees Based On Logistic Regression

27

30% Loss In Accuracy Occurs When Explaining Using Decision Trees

0% Loss With Adaptive Insight Using Illuminated AI

(In FinTech, Medicine, Government, …)

Histogram of Factors that Hinder vs. Help the Case

Case #1 Score 0.75 Not Approved Case #2 Score 0.81 Borderline Approved Case #3 Score 0.98 Strongly Approved

Understanding Decisions at a Glance

Number of Factors

• 0.25 -0.2 -0.15 -0.1 -0.05 0

0.05 0.1 0.15 2 4 6 8 10 12

Helps Hinders

Most Hindering Factor: E With value 66.2 26.8 below expected

2 4 6 8 10 12 14

Helps Hinders

• 0.5

0.5 1 1.5

Most Helping Factor: G With value -8.8 5.8 above expected

0.6

• 0.2 -0.1

0.1 0.2 0.3 0.4 0.5 5 10 15 20 25

Helps Hinders

Most Helping Factor : M with score 87.0 12.2 above expected

28

Comparison:

Structure (during recognition)

Feedforward Illuminated

“Feedforward”

Outputs Y Inputs X

Y X

W

Feedforward-Feedback

Outputs Y Inputs X

Y X

M

M

Explainable?

Yes No

Easy to Learn & Update?

Yes No

Optimization

During Learning During Recognition

29

Collective Benefits Enabling Wider AI Adoption

Company

Reduce: development costs, time Increase: trust, adoption, and flexibility

Users

Better understanding and trust of AI’s decision process

Developers

Less guessing, easier debugging and updating

Regulators

Understanding of AI’s goals, compromises, and decision process

30

Offerings

• Convert & Explain Any Feedforward AI
• Boost: Internal Development & Quality Assurance
• Assist Your Regulators: FDA, AMA, DMV, FTC …

1) Illuminate Your AI

• Faster Learning - 100x
• Less Data Cleaning
• User Personalization
• Easier Update

* for Certain Feedforward AI

2) Train or Update Your AI Fast Without Rehearsal*

achler@OptimizingMind.com

How Will You Use Illuminated AI?

32

Tsvi Achler 650.486.2303