The human intellect is like peacock feathers, just an extravagant display intended to attract a mate. Oh we think we are so great ..aha.. but a peacock can barely fly and eats insects.

What is Vertical AI?

The Big Nine have open-sourced their toolkits for horizontal AI

Company Product

DLKIT

CNTK

DSSTNE

Citizen doctors

We now have supercomputers with thousands of cores (GPUs)3584*4=14,336 CUDA cores , $1/core

sentence = """This is what I have to you ..DL approach to NLP.""" tokens = nltk.word_tokenize(sentence)tagged = nltk.pos_tag(tokens)entities = nltk.chunk.ne_chunk(tagged)from nltk.corpus import treebank t = treebank.parsed_sents('wsj_0001.mrg')[0]t.draw()as alternative ... vocabulary_size = 8000 unknown_token = "UNKNOWN_TOKEN" sentence_start_token = "SENTENCE_START" sentence_end_token = "SENTENCE_END"# Read the data and append SENTENCE_START and SENTENCE_END tokensprint "Reading my CV file..."with open('CV-EN_v0.6.csv', 'rb') as f:reader = csv.reader(f, skipinitialspace=True)reader.next() # Split full comments into sentencessentences = itertools.chain(*[nltk.sent_tokenize(x[0].decode('utf-8').lower()) for x in reader])# Append SENTENCE_START and SENTENCE_ENDsentences = ["%s %s %s" % (sentence_start_token, x, sentence_end_token) for x in sentences] print "Parsed %d sentences." % (len(sentences)) # Tokenize the sentences into wordstokenized_sentences = [nltk.word_tokenize(sent) for sent in sentences]# Count the word frequenciesword_freq = nltk.FreqDist(itertools.chain(*tokenized_sentences))print "Found %d unique words tokens." % len(word_freq.items())# Get the most common words and build index_to_word and word_to_index vectorsvocab = word_freq.most_common(vocabulary_size-1)index_to_word = [x[0] for x in vocab]index_to_word.append(unknown_token) word_to_index = dict([(w,i) for i,w in enumerate(index_to_word)])print "Using vocabulary size %d." % vocabulary_sizeprint "The least frequent word in our vocabulary is '%s' and appeared %d times." % (vocab[-1][0], vocab[-1][1]) # Replace all words not in our vocabulary with the unknown tokenfor i, sent in enumerate(tokenized_sentences):tokenized_sentences[i] = [w if w in word_to_index else unknown_token for w in sent]print "\nExample sentence: '%s'" % sentences[0]print "\nExample sentence after Pre-processing: '%s'" % tokenized_sentences[0]# Create the training dataX_train = np.asarray([[word_to_index[w] for w in sent[:-1]] for sent in tokenized_sentences])y_train = np.asarray([[word_to_index[w] for w in sent[1:]] for sent in tokenized_sentences])

do i have your attention now?

We are DATAWE ARE CODE

Government is aligned

The Chinese Academy of Sciences has issued invitations to apply for

funding for PMI projects worth

$9.2 bn

NIH is funding PMI projects worth

$215 M

MEDICAL RECORD: The unfair fight for what is ours

IHTFPIHTFP

SCRA

PE50% Self-reported information20% biomarkers (blood, urine results)

10% scans20% doctor notes

REC

REAT

E70%

30%

Everything that does not learn will disappear

DIALOG SYSTEMS in a CLOSED DOMAIN

VA⊃BOT Supervised Learning

Discriminative Models: CRF, RF, NN, SVM SA, state tracker

Logistic regression for intentions (analysis) (CRF) labels + script answers (generation)

4.0 Artificial Intelligence(AI) her

Rule-based Pattern matching

PyAIML

• 3.0 Intelligent Agents (IA) doc

• 2.0 Virtual Assistant (VA) Amy, Clara, Melody

• 1.0 Chatbot (BOT) Mitsuku, Alice, Botlibre, Luka

AI⊇IA Zeno Behavior : 0.9999999..

IA⊇VA

Learning methodologies 1. Unsupervised (clustered patterns, machine generated) 2. Supervised (labels, classification) 3. RL (by experts, pre-release) 4. Memory-based (where concrete cases are used at runtime, so

learning is "just" remembering) 5. Semantic embedding: word2vec 6. “One-shot learning”

Algorithms

‣ clustering (k-NN) ‣ classification (discrete) ‣ regression (continuous) ‣ structural matching to match present circumstance in past

cases (CB) ‣ structural elaboration and substitution (use past examples for context-appropriate generation)

1. faceted (allowing users to explore a collection of information by applying multiple filters)

2. autocomplete 3. multimode 4. Speaker analysis 5. Goal-oriented dialogue 6. Longitudinal representation (Micro-Moments which are intensive-rich)

CONVERSATIONAL SEARCH

Linguistic Authority

1. NLU (bAbi, SemEval) or Winograd Schema = Google-proof PDP

2. MemNN 3. bidirectional CG (RRG, TAG, FCG, HPSG, CCG, DG,

LFG..) 4. OFSM (Optimizing Finite State Machines)

KR/Ontology ‣ LOINC (clinical values) ‣ ICD9, ICD10 (diagnoses) ‣ CPT (medical services) ‣ SNOMED (EHR terms) ‣ RXNORM (medications ‣ MeSH (how to combine structured vocabularies with data-driven

techniques)

Memory ‣ MemNN ‣ app ‣ FSM ‣ Pontifex Maximus

Association mining: Location-aware (Item-item CF recommender using DSSTNE)

IHTFP

RESUME

Name: doc Nationality: AI Age:1.577e+10ms Weight: 28.085(atomic 28.084–28.086) Address: Exxact Spectrum TXR410-0032R Family: NVIDIA DIGITS™ IQ: 14,336 CUDA cores Atomic number: 14 Gender: Si Student: big data medicine University: Enigma.ai Profession: rob0-medstudent Drivers license:4x NVIDIA GTX 1070/1080 Pascal GPUs Character: stochastic Language: python,C++, LUA, English, Mandarin Graduation: 2017 Ambition in life: helping my carbon-based medical colleagues with number crunching and their students (patients) with education Courses: crunching of genomes and phenomes Semantic neurons:LOINC, ICD9, ICD10, CPT, SNOMED, RXNORM Mission: Medical Access 4 all Thesis: n=1 Hobbies: TensorFlow, Torch

You are todayI am tomorrow

CORTEX

Top down: Model the Target

medication/therapy

disease/diagnosis

results

scans/blood test

human

medical record

symptom

Discriminate models(supervised learning)• SVM• NN• RF• CRF

Bottom up: Target the Model

OMICS

Genome

Health

Blood panel

Phenome

Blood

Generative Models(unsupervised learning)• HMM• Naive Bayes• DNN• RBM

Data-driven models

Conversational Intelligent Agents (CIAs) are a BLUE OCEAN market space.

LicensingLead Generation

Store

B2B

sponsored bots

Native Content

Affiliate Marketing

Surveys

ConsultationsTherapy

$200Bn industry

GENOMICS

1 MILLION GENOMES (2016)

1 BILLION GENOMES (2023)

World Shortage of doctors: 5,000,000

CRISPR will boost Personal Genomics

Data becomes a revenue stream

Health: A Novel Asset Category

LIFE EXPECTANCY QUARTILE CREDIT TRANCHE INSURANCE HEALTHCARE

73.9

82.32

89.31

95-100

LOWER

MEDIAN

UPPER

LEVERAGED

AA-

AA

AA+

AAA

JUNIOR

SENIOR

SUPER SENIOR

LSS

STANDARD

SELECT

PREFERRED

PREFERRED PLUS

REACTIVE

ACTIVE

PROACTIVE

LIFE EXTENSION

AIRLINES

ECONOMY

ECONOMY+

BUSINESS

FIRST CLASS

∆= 21st C20th

n = 1

21st C20th

N x*

Healthcare will become a continuous function

Discrete function

Continuous function

More is More

.56=0.015625

.512=0.000244140625

walter@doc.ai

@walterdebrouwer

walterdebrouwer

www.doc.ai

walterdb

doc Incorporated 540 Bryant Street Palo Alto, CA 94301