Advancing the Future of AI IBM Research AI February 2021 Introduction Trusted AI Scaling and Automating AI

Dr. John R. Smith, IBM Fellow Dr. Karthikeyan N. Ramamurthy Dr. Lisa Amini, Director IBM T. J. Watson Research Center IBM T. J. Watson Research Center IBM Research Cambridge

2 Introduction

Dr. John R. Smith, IBM Fellow IBM T. J. Watson Research Center

3 IBM Research AI AI Focus Areas include: • Global footprint with an integrated AI strategy Bot • Network of academic partnerships in AI

Customer Agent

Language Speech Code

b c

a d

Time series Hardware Chemistry

Dublin ... more Zurich Yorktown Cambridge Almaden Tokyo Haifa Delhi

Bangalore

MIT - IBM

A I L A B Narrow AI Broad AI General AI Emerging Disruptive and pervasive Revolutionary

Deep learning Neuro-symbolic AI True neuro-AI

Single-task, Trusted AI capable of learning Cross-domain learning single-domain, with much less data and reasoning high accuracy Automating AI development Broad autonomy with Requires large and deployment moral reasoning amounts of data Reduced-precision and Wetware? CPU and GPU analog HW

We are here now AI is making incredible impact

Example Applications (Accelerated discovery) What’s Next (Advances in AI foundations)

Knowledge Ingestion Neuro-symbolic and Reasoning Reasoning www.research.ibm.com/artificial- www.research.ibm.com/covid19/deep-search/ intelligence/vision/#neurosymbolic

Deep Search (PDF Corpora) Logical Neurons

Chemical Reaction Trusted AI Prediction www.research.ibm.com/artificial- intelligence/trusted-ai/ https://rxn.res.ibm.com Uncertainty Quantification RXN for Chemistry

Cloud-based Auto AI Autonomous Labs https://www.ibm.com/cloud/watson- studio/autoai https://rxn.res.ibm.com/rxn/robo-rxn RoboRXN (AI-Driven Synthesis) AutoAI

January 2021 IBM Research – Advancing AI for Science 6 Trusted AI

Dr. Karthikeyan N. Ramamurthy IBM T. J. Watson Research Center

7 AI is powering critical workflows and trust is essential

loan customer employment quality control processing management

brand reputation increased regulation complexity of AI deployments focus on social justice

is it fair? is it accurate? did anyone tamper with it?

is it easy to understand?

is it accountable? is it transparent? 10

IBM Research / February 21, 2021 / © 2021 IBM Corporation “A decision aid, no matter how sophisticated or ‘intelligent’ it may be, may be rejected by a decision maker who does not trust it, and so its potential benefits to system performance will be lost.” —Bonnie M. Muir, psychologist at University of Toronto

IBM Research / February 21, 2021 / © 2021 IBM Corporation “The toughest thing about the power of trust is that it’s very difficult to build and very easy to destroy.” —Thomas J. Watson, Sr., CEO of IBM

Verify

problem Take advice from a panel owner

diverse problem data of diverse voices stakeholders specification understanding data engineer

data preparation

deployment and monitoring

AI modeling operations engineer data scientist

evaluation

model validator 15

data preparation measurement sampling construct space observed space raw data prepared data space space

data preparation social bias population bias bias

temporal bias data poisoning

threatens

threatens threatens

internal construct validity external validity validity

initial model final model pre-processing pre-processed model training post-processing training dataset dataset

general robustness adversarial robustness fairness explainability

initial model final model pre-processing pre-processed model training post-processing training dataset dataset

data augmentation invariant risk minimization post hoc correction data sanitization gradient shaping post hoc defense bias mitigation pre-processing bias mitigation in-processing bias mitigation post-processing disentangled representations directly interpretable models post hoc explanations

Consulting

– Trust capabilities in IBM Cloud Pak for Data • Explainability Fundamental Early • Fairness Research Access • Drift detection – Open source toolkits • AI Explainabilty 360 • AI Fairness 360 Open IBM Cloud Source Pak for Data • Adversarial Robustness Toolbox – Early access to enhanced editions of toolkits

AI FactSheets 360 consulting human-computer interaction

lending transparency specification elicitation

transparent documentation and eliciting societal values and preferences from policymakers are critical for AI governance workforce FACT DART FreaAI VerifAI Uncertainty Quantification 360

testing uncertainty quantification

health domain domain expertise

all elements of trust should be tested and reported with error bars

AI Explainability 360 AI Fairness 360 Adversarial Robustness 360 Trusted Generation 360 these are social issues too issuessocial are these

ecosystem explainability fairness robustness generation

causality and constraint-aware learning are important in their own right and also as foundations for the pillars of trust above

topics and assets should not be viewed only through a technical lens; lens; technical a through only viewed be not should assets and topics Causallib Diffprivlib broader broader impacts government affairs government causal modeling constraint-aware learning

Use master pages to ensure – With the variety of layout solutions, one or more layouts should work to showcase your content. consistency throughout your – Using this template will help avoid overly dense presentation. pages and fonts that are unreadable. This will present a more unified look across presentations. – Master layouts are grouped by background color: black and light gray. – Each master group includes 32 page designs for you to choose from.

Supporting diverse and rich – 8 unique techniques from IBM Research explanations. • data vs. model • global vs. local • direct vs. post hoc http://github.com/ibm/aix360 – LIME and SHAP – 2 explainability metrics – Extensive industry tutorials to educate users and http://aix360.mybluemix.net practitioners – Interactive demo

Early access offering – New state-of-the-art explanation methods from IBM Research – Interactive explanation – Better support for text – Better support for regression

The most comprehensive – Comprehensive set of fairness metrics toolkit for handling bias in • Group fairness machine learning. • Individual fairness – 12 state-of-the-art bias mitigation algorithms • Pre-processing http://github.com/ibm/aif360 • In-processing • Post-processing – Extensive industry tutorials to educate users and http://aif360.mybluemix.net practitioners – Interactive demo

Early access offering – New state-of-the-art bias mitigation algorithms from IBM Research and the external community – Support for natural language processing – Bias mitigation and verification for individual fairness – Support for transfer learning – Fair generation – Protected attribute extraction

problem Take advice from a panel owner

diverse problem data of diverse voices stakeholders specification understanding data engineer

data preparation

deployment and monitoring

AI modeling operations engineer data scientist

evaluation

model validator 29

Predictive Performance Data Scientist Model Validator AI Ops Engineer AI Factsheets 360 - https://aifs360.mybluemix.net/ Recap of the Toolkits

AI Fairness 360 AI Explainability 360 Adversarial Robustness 360 FactSheets 360 aif360.mybluemix.net aix360.mybluemix.net art-demo.mybluemix.net aifs360.mybluemix.net

Most comprehensive open source Most comprehensive open source Most comprehensive open source Extensive website describing toolkit for detecting & mitigating toolkit for explaining ML models & toolkit for defending AI from attacks research effort to foster trust in AI bias in ML models: data • Supports 10+ frameworks by increasing transparency and • 70+ fairness metrics • 8 explainability algorithms • 19 composable and modular attacks enabling • 12 bias mitigators • Interactive demo showing 3 algorithms (including adaptive white- and black- Governance box) • Interactive demo illustrating 5 bias in credit scoring application • 6 examples FactSheets • 10 defenses, including detection of metrics and 4 bias mitigators • 13 tutorial notebooks: finance, • 7-step methodology for creating adversarial samples and poisoning • extensive industry tutorials and healthcare, lifestyle, retention, etc. useful FactSheets attacks notebooks • Extensive documentation and • AI Governance • Robustness metrics, certifications and taxonomy of explainability algorithms • Papers, videos, related work, FAQ, verifications slack channel • 30 notebooks covering attacks and defenses • From dozens of publications Scaling and Automating AI

Dr. Lisa Amini, Director IBM Research Cambridge

33 AI Agenda for the Enterprise

Advancing AI core algorithms for mastering language, learning and Advancing AI reasoning From narrow to broad AI Trusting AI through fairness, robustness, explainability, and Trusting AI transparency

Operationalizing Scaling AI by managing, operating & Scaling AI AI at scale automating its lifecycle Scaling and automating AI A holistic approach

Data Automation ML & Data Science Automation

Data Pipeline- Quality Cleaning & Feature Model train& Quality specific Selection & Labeling Augmentation Modeling Analysis Prep Creation test Assurance Guardrail Gap Analysis Definition

Auto ML (automated machine learning)

Monitor/Adapt Production Operations

Test, Analyze/ Refit on Deploy Monitor Validation Recommend Additional Data

IBM Confidential | © 2021 IBM Corporation AI Lifecycle and Governance Automation Readiness Validators Guide user in understanding reporting data readiness for ML model Data Remediators building, and mitigating issues Generation

AI-assisted Constraints Data Readiness Toolkit (DaRT) Labeling profiling

IBM Research’s Framework for Federated Learning (FFL) and Automation for data Aggregator advanced Model Fusion algorithms enable training

Party A Party n models across disparate systems, keeping data in place

Party B

Automate feature engineering for relational data with complex Prepare Create, Train Validate & Monitor & connections Data Model Deploy Adapt OneButtonMachine

Knowledge Augmentation technology automatically augments your data with domain knowledge from notebooks, documents,

Existing code developed Documents containing External knowledge by data scientists domain knowledge

Knowledge Graphs DBpedia, Wikidata, YAGO Mathematical/ Engineering/ Business Domain Open Data e.g. data.gov Documents containing a mixture of natural Web Tables HTML tables on the Web language and mathematical formulae Blogs, newspapers e.g. NY Times

NAME AGE INCOME CREDIT NAME AGE INCOME CREDIT Loan / Income Hana Ja 34 $50,000 $20,000 Hana Ja 34 $50,000 $20,000 0.4 Taro Tok 25 $40,000 $20,000 Taro Tok 25 $40,000 $20,000 0.5 Input data to ML Problem Augmented data

KAFE: Automated Feature Enhancement for Predictive Modeling using External Knowledge Sainyam Galhotra, Udayan Khurana, Oktie Hassanzadeh, Kavitha Srinivas and Horst Samulowitz KR2ML Workshop at NeurIPS, 2019 IBM Confidential | © 2021 IBM Corporation Automatically augmenting data with knowledge

2 Feature-concept Mapping

3 Expression Extraction

Natural Language Documents (i.e. Wikipedia, internal documents) 1 Knowledge Base

1 Leverage Knowledge Bases Merging, unifying and smart indexing of existing knowledge bases 2 Feature to concept mapping Map data and features to concepts (eg., numbers to ‘income’) 3 Expression extraction Extract formulas and expressions (such as ‘income divided by loan amount’) 4 Feature construction Compute novel features based on concept mapping and expressions Automation for more complex modalities & tasks

Prepare Create, Train Validate & Monitor & Data Model Deploy Adapt

AI operations – Cloud budget < $1 million – Timeline: 4 weeks

Compliance – Minimize bias – Provide explanations for approval/denial – Race can never be used as a feature

Application developer – Model size < 1GB Data scientist – Time per prediction < 30ms

Bank wants to determine an appropriate interest rate for a new loan IT operations – Given high dimensions of data use Task: Predict credit risk for an application support vector machines – Apply previously generated models

Alternating direction method of multipliers (ADMM) based algorithm solves this complex optimization problem by splitting into easier subproblems.

Black-box Constraints Method Selection Hyper-Parameter Optimization

– Deployment constraints – Combinatorial multi-armed bandits – Random search (can be competitive) – Fairness – Genetic algorithms – Sequential Model-Based Optimization – Explainability & Interpretabiliity – Multi-fidelity approximation – Gaussian Processes – Problem & domain – Trust-region DFO – Physical & natural

Formulation enables automation well beyond ML model building

Prepare Create, Train Validate & Monitor & Data Model Deploy Adapt

Pre-Processing Algorithms In-Processing Algorithms Post-Processing Algorithms Mitigates Bias in Training Data Mitigates Bias in Classifiers Mitigates Bias in Predictions

Reweighing Meta Fair Classifier Rejection Option Classification Modifies the weights of different training Takes fairness metric as part of input & Changes predictions from a classifier to examples returns classifier optimized for the metric make them fairer

Disparate Impact Remover Prejudice Remover Calibrated Equalized Odds Edits feature values to improve group Adds a discrimination-aware Optimizes over calibrated classifier score fairness regularization term to learning objective outputs that lead to fair output labels

Optimized Preprocessing Adversarial Debiasing Equalized Odds Modifies training data features and labels Uses adversarial methods to maximize Modifies the predicted label using an accuracy & reduce evidence of protected optimization scheme to make predictions attributes in predictions fairer Learning Fair Representations Learns fair representations by obfuscating information about protected values

Prepare Create, Validate & Monitor & Data Train Model Deploy Adapt

https://aif360.mybluemix.net How do Data Science Workers Collaborate?: Roles, Workflows, and Tools, CSCW, 2020. How Data Scientists Work Together with Domain Experts in Scientific Collaborations: To Find the Right Answer or To Ask the Right Question?, GROUP, 2020. Time series forecast, scaling, and automation

Auto generate forecasting Support large scale time Handle irregularly sampled Enable augmentation with pipelines with algorithms: series data: data: external variables:

– Auto feature engineering and Typical data sets from many – Data isn’t always well behaved Signals from weather and other hyperparameter optimization use cases may contain: (uniformly sampled, with no data sources can help the missing values). prediction process for the target tailored for time series – 1000s of time series time series. – Conventional approaches for – Support multi-series and multi- – 100k+ data points per series step forecasting resampling or imputation may Handling this scale requires not yield good results. – Configurable back testing special architectures and – Leverage state of the art deep setup techniques. learning models to address these challenges.

Historical Historical Business Decisions Impact

Demand forecasting Additional information to model decision Automatically generate model based on history from “Standard AutoAI” making from historical data (e.g. using regression, state/action reward)

– Many patterns amenable to data-driven automation: Set point optimization, Continuous Flow Maximization, Inverse Optimization, RL for sequential decision optimization. – Enables data-driven learning of risk-measures, scenarios, or uncertainty sets

Training Stronger Baselines for Learning to Optimize, NeurIPS 2020 Watson AutoAI: Can I get the model? Medium.com https://pages.github.ibm.com/factsheets/ai-gov-demo/ https://www.youtube.com/watch?v=-4d3kEVsu-s&t=2s AI lifecycle automation

Prepare Create, Train Validate & Monitor & Data Model Deploy Adapt

Scores Confidence cat .72 Labeled dog .95 1. Train the model Test Data dog .63 Model cat .51 2. Score the test set with the model and evaluate 3. Construct new training data for the Meta-model Additional meta-model from the model’s Features Labels training data features scores on the test data 4. Train the meta-model Additional features 5. Score production data with the meta-model to obtain accuracy Scores Unlabeled predictions production log data Performance Prediction Meta Model

Problem I need to introduce a new version of a model into my application with minimal risk Over time, the bandit learns to shift towards most rewarding Safe deploy model version Our Approach Staged deployment with Contextual Bandits that transparently deploys new versions safely without interrupting the application

Bandit explores available model versions to learn

Prepare Create, Validate & Monitor & Data Train Model Deploy Adapt

Composable capabilities across the AI lifecycle Orchestrated through reusable pipelines

Data Selection, Quality, Management, Labeling Trust Insight, Analysis & Improvement

DaRT Data Active Conflict Performance Business KPI Improvement AIF 360 AIX 360 Readiness Learning Analysis Prediction Insight Recommender Toolkit Toolkit Toolkit Toolkit Toolkit Toolkit Toolkit Toolkit

New Class Reward Based AI assisted AI Robustness Detection Data Labeling Learning Unsupervised Toolkit Toolkit Toolkit Toolkit Learning

Simplifying and accelerating the entire, end-to-end AI Lifecycle

Continuously improving your AI models and applications

Expanding our scope of Automation

AutoAI with Data Quality and Business Constraints Knowledge Augmentation for Supervised Learning: – AutoAI Playground: https://autoai.mybluemix.net/ – KAFE: Automated Feature Enhancement for Predictive Modeling using External – Lale Git repo: https://github.com/IBM/lale/ Knowledge – Lale: Consistent Automated Machine Learning. Sainyam Galhotra, Udayan Khurana, Oktie Hassanzadeh, Kavitha Srinivas and Horst Guillaume Baudart, et. al., KDD Workshop AutoML@KDD, 2020. Samulowitz https://arxiv.org/abs/2007.01977 KR2ML Workshop at NeurIPS, 2019 – Mining Documentation to Extract Hyperparameter Schemas. – Automated Feature Enhancement for Predictive Modeling using External Knowledge Guillaume Baudart, et. al. ICML Workshop AutoML@ICML, 2020. Sainyam Galhotra, Udayan Khurana, Oktie Hassanzadeh, Kavitha Srinivas, Horst https://arxiv.org/abs/2006.16984 Samulowitz, Miao Qi – Solving Constrained CASH Problems with ADMM. IEEE ICDM (demo track), 2019 Parikshit Ram, et al. ICML Workshop AutoML@ICML, 2020. – Semantic Search over Structured Data, Sainyam Galhotra and Udayan – Data Quality for AI: https://w3.ibm.com/w3publisher/ibm-research-ai-business- Khurana [Nominated for Best Paper CIKM2020] development/ai-portfolio/data-quality-for-ai https://dl.acm.org/doi/10.1145/3340531.3417426 – An ADMM Based Framework for AutoML Pipeline Configuration. Liu, Sijia, et al. AAAI. 2020. Auto AI Lifecycle for Time Series: – Human-AI Collaboration in Data Science: Exploring Data Scientists' Perceptions of – FLOps: On Learning Important Time Series Features for Real-Valued Prediction. Automated AI. Wang, Dakuo, et al. CSCW (2019) Dhaval Patel, et. al., IEEE Big Data 2020 – AutoAI: Automating the End-to-End AI Lifecycle with Humans-in-the-Loop. – Smart-ML: A System for Machine Learning Model Exploration using Pipeline Graph. Dakuo Wang, et al. ICIUI, 2020. Dhaval Patel, et. al.,. IEEE Big Data 2020 – AutoAIViz: opening the blackbox of automated artificial intelligence with conditional – ThunderML: A Toolkit for Enabling AI/ML Models on Cloud for Industry 4.0S parallel coordinates. Daniel Weidele, et al. ICIUI, 2020. Shrivastava, D Patel, WM Gifford, S Siegel, J Kalagnanam, International Conference on – Selecting Near-Optimal Learners via Incremental Data Allocation Web Services, 163-180 Ashish Sabharwal, Horst Samulowitz and Gerald Tesauro. AAAI 2016. – Providing Cooperative Data Analytics for Real Applications Using Machine Learning, A Iyengar, et. al. IEEE ICDCS, 2019 Scaling AI through Federated Learning and Fast Inference – Model Agnostic Contrastive Explanations for Structured Data. – IBM federated learning Git repo: https://github.com/IBM/federated-learning-lib Dhurandhar et al. – IBM federated learning web page: https://ibmfl.mybluemix.net – Boolean decision rules via column generation. – IBM federated learning White paper: https://arxiv.org/abs/2007.10987 Dash, Sanjeeb, Oktay Gunluk, and Dennis Wei. NeurIPS 2018. – Snap ML project web page: https://www.zurich.ibm.com/snapml/ – Snap ML documentation: https://ibmsoe.github.io/snap-ml-doc/ IBM AutoAI Product Resources – Snap ML NeurIPS 2020 paper: https://arxiv.org/abs/2006.09745 – IBM Developer: https://developer.ibm.com/series/explore-autoai/ – Coursera course: https://www.coursera.org/learn/ibm-rapid-prototyping-watson- studio-autoai – AutoAI demo: https://www.ibm.com/demos/collection/IBM-Watson-Studio-AutoAI/ IBM Confidential | © 2021 IBM Corporation