panel incremental learning: how systems can ... · els lefever: sentimentanalysis 17...
TRANSCRIPT
![Page 1: PANEL INCREMENTAL LEARNING: HOW SYSTEMS CAN ... · ELS LEFEVER: SENTIMENTANALYSIS 17 Domain-adaptation: retrain on domain-specific corpus (e.g. “predictable” for a movie plot](https://reader035.vdocuments.mx/reader035/viewer/2022070710/5ec742453461017ab76b6522/html5/thumbnails/1.jpg)
PANEL
INCREMENTAL LEARNING:
HOW SYSTEMS CAN AUTOMATICALLY ADAPT TO
AND LEARN FROM NEW SITUATIONS
moderator: Els Lefever
![Page 2: PANEL INCREMENTAL LEARNING: HOW SYSTEMS CAN ... · ELS LEFEVER: SENTIMENTANALYSIS 17 Domain-adaptation: retrain on domain-specific corpus (e.g. “predictable” for a movie plot](https://reader035.vdocuments.mx/reader035/viewer/2022070710/5ec742453461017ab76b6522/html5/thumbnails/2.jpg)
Incremental learning = a “machine learning paradigm where the learningprocess takes place whenever new example(s) emerge and adjusts whathas been learned according to the new example(s)” (Geng & Smith-Miles,2015)
Traditional machine leaning: implicit assumption that a “good” trainingset in a domain is available a priori => the training set contains allnecessary knowledge that once learned, can be reliably applied to anynew examples in the domain
In practice: many real-world applications cannot match this ideal case
2
![Page 3: PANEL INCREMENTAL LEARNING: HOW SYSTEMS CAN ... · ELS LEFEVER: SENTIMENTANALYSIS 17 Domain-adaptation: retrain on domain-specific corpus (e.g. “predictable” for a movie plot](https://reader035.vdocuments.mx/reader035/viewer/2022070710/5ec742453461017ab76b6522/html5/thumbnails/3.jpg)
, University of Houston, USA:“Closed Loop Deep Brain Stimulation in Movement Disorders”
, Brandeis University, USA:“Using artificial neural networks to analyze biological neurons via
transfer learning”
, Ort Braude College, Israel:“Incremental reasoning on strongly distributed systems”
, Ghent University, Belgium:“domain/context-specific and implicit sentiment analysis”
3
![Page 4: PANEL INCREMENTAL LEARNING: HOW SYSTEMS CAN ... · ELS LEFEVER: SENTIMENTANALYSIS 17 Domain-adaptation: retrain on domain-specific corpus (e.g. “predictable” for a movie plot](https://reader035.vdocuments.mx/reader035/viewer/2022070710/5ec742453461017ab76b6522/html5/thumbnails/4.jpg)
ELENA RAVVEORT BRAUDE COLLEGE, ISRAEL
![Page 5: PANEL INCREMENTAL LEARNING: HOW SYSTEMS CAN ... · ELS LEFEVER: SENTIMENTANALYSIS 17 Domain-adaptation: retrain on domain-specific corpus (e.g. “predictable” for a movie plot](https://reader035.vdocuments.mx/reader035/viewer/2022070710/5ec742453461017ab76b6522/html5/thumbnails/5.jpg)
5
![Page 6: PANEL INCREMENTAL LEARNING: HOW SYSTEMS CAN ... · ELS LEFEVER: SENTIMENTANALYSIS 17 Domain-adaptation: retrain on domain-specific corpus (e.g. “predictable” for a movie plot](https://reader035.vdocuments.mx/reader035/viewer/2022070710/5ec742453461017ab76b6522/html5/thumbnails/6.jpg)
6
![Page 7: PANEL INCREMENTAL LEARNING: HOW SYSTEMS CAN ... · ELS LEFEVER: SENTIMENTANALYSIS 17 Domain-adaptation: retrain on domain-specific corpus (e.g. “predictable” for a movie plot](https://reader035.vdocuments.mx/reader035/viewer/2022070710/5ec742453461017ab76b6522/html5/thumbnails/7.jpg)
PENGYU HONGBRANDEIS UNIVERSITY, USA
![Page 8: PANEL INCREMENTAL LEARNING: HOW SYSTEMS CAN ... · ELS LEFEVER: SENTIMENTANALYSIS 17 Domain-adaptation: retrain on domain-specific corpus (e.g. “predictable” for a movie plot](https://reader035.vdocuments.mx/reader035/viewer/2022070710/5ec742453461017ab76b6522/html5/thumbnails/8.jpg)
8
Learning
Task: Train Artificial Neural Networks to
Analyze Artificial Neural Networks
![Page 9: PANEL INCREMENTAL LEARNING: HOW SYSTEMS CAN ... · ELS LEFEVER: SENTIMENTANALYSIS 17 Domain-adaptation: retrain on domain-specific corpus (e.g. “predictable” for a movie plot](https://reader035.vdocuments.mx/reader035/viewer/2022070710/5ec742453461017ab76b6522/html5/thumbnails/9.jpg)
EFFECTIVE LEARNING
9
Deep Learning from Large Amount of Synthetic Data andSmall Amount of Labeled Real Data
Apply to
Learn from “Toy” Examples and Generalize to Real Scenarios
![Page 10: PANEL INCREMENTAL LEARNING: HOW SYSTEMS CAN ... · ELS LEFEVER: SENTIMENTANALYSIS 17 Domain-adaptation: retrain on domain-specific corpus (e.g. “predictable” for a movie plot](https://reader035.vdocuments.mx/reader035/viewer/2022070710/5ec742453461017ab76b6522/html5/thumbnails/10.jpg)
NURI INCEUNIVERSITY OF HOUSTON, USA
![Page 11: PANEL INCREMENTAL LEARNING: HOW SYSTEMS CAN ... · ELS LEFEVER: SENTIMENTANALYSIS 17 Domain-adaptation: retrain on domain-specific corpus (e.g. “predictable” for a movie plot](https://reader035.vdocuments.mx/reader035/viewer/2022070710/5ec742453461017ab76b6522/html5/thumbnails/11.jpg)
Functional Use Of Local Field Potentialsfor the Personalization of DBS in
Parkinson’s Disease
Nuri F. Ince, Ph.D.Clinical Neural Engineering Laboratory (CNELab)
Department of Biomedical Engineering, University of Houston
12/25/2016 11
PosturalInstability
Tremor
Rigidity
Bradykinesia
![Page 12: PANEL INCREMENTAL LEARNING: HOW SYSTEMS CAN ... · ELS LEFEVER: SENTIMENTANALYSIS 17 Domain-adaptation: retrain on domain-specific corpus (e.g. “predictable” for a movie plot](https://reader035.vdocuments.mx/reader035/viewer/2022070710/5ec742453461017ab76b6522/html5/thumbnails/12.jpg)
PARKINSON’S DISEASE• Parkinson's disease (PD) was first described in 1817 by
James Parkinson as a particular form of progressive motordisability (Samii, Nutt, & Ransom, 2004).
• PD is a chronic and progressive movement disorder,meaning that symptoms continue and worsen over time.
• Today, PD is the second most common neurodegenerativedisorder after Alzheimer’s dementia.
• Nearly one million people in the US are living with PD.
• Although there is presently no cure, there are treatmentoptions such as medication and DBS to manage itssymptoms.
• DBS very effective but
• Stim side effects
• Open Loop
• Does not adapt fluctuating symptoms
• Battery runs out in 3-5 years
ELECTRICAL STIMULATION
![Page 13: PANEL INCREMENTAL LEARNING: HOW SYSTEMS CAN ... · ELS LEFEVER: SENTIMENTANALYSIS 17 Domain-adaptation: retrain on domain-specific corpus (e.g. “predictable” for a movie plot](https://reader035.vdocuments.mx/reader035/viewer/2022070710/5ec742453461017ab76b6522/html5/thumbnails/13.jpg)
BIDIRECTIONAL INTERFACE
Local Field Potentials
• Do they need any adaptation?
• Can we detect fluctuations in symptoms?
• Can we detect states such as sleep, awake?
• Can we tune stimulation for each patient?
• Can we extend battery life?
• Can we integrate sensors with neurophysiology?
Activa PC+S
![Page 14: PANEL INCREMENTAL LEARNING: HOW SYSTEMS CAN ... · ELS LEFEVER: SENTIMENTANALYSIS 17 Domain-adaptation: retrain on domain-specific corpus (e.g. “predictable” for a movie plot](https://reader035.vdocuments.mx/reader035/viewer/2022070710/5ec742453461017ab76b6522/html5/thumbnails/14.jpg)
ELS LEFEVERGHENT UNIVERSITY, BELGIUM
![Page 15: PANEL INCREMENTAL LEARNING: HOW SYSTEMS CAN ... · ELS LEFEVER: SENTIMENTANALYSIS 17 Domain-adaptation: retrain on domain-specific corpus (e.g. “predictable” for a movie plot](https://reader035.vdocuments.mx/reader035/viewer/2022070710/5ec742453461017ab76b6522/html5/thumbnails/15.jpg)
Sentiment analysis (or opinion mining) = the“computational treatment of opinion, sentiment, andsubjectivity in text” (Pang & Lee, 2008)
15
![Page 16: PANEL INCREMENTAL LEARNING: HOW SYSTEMS CAN ... · ELS LEFEVER: SENTIMENTANALYSIS 17 Domain-adaptation: retrain on domain-specific corpus (e.g. “predictable” for a movie plot](https://reader035.vdocuments.mx/reader035/viewer/2022070710/5ec742453461017ab76b6522/html5/thumbnails/16.jpg)
ELS LEFEVER: SENTIMENT ANALYSIS
16
Friendly staff,
Very quiet location,
bit remote though.
Friendly
quiet
remote
great
Success
nightmare
mistake
horrible
Machine learning approaches topredict polarity: “positive”, “negative” or“neutral” label
Incremental learning: system is
designed as a semi-supervised
intelligent system, which improves
in each iteration as we get more
training data
![Page 17: PANEL INCREMENTAL LEARNING: HOW SYSTEMS CAN ... · ELS LEFEVER: SENTIMENTANALYSIS 17 Domain-adaptation: retrain on domain-specific corpus (e.g. “predictable” for a movie plot](https://reader035.vdocuments.mx/reader035/viewer/2022070710/5ec742453461017ab76b6522/html5/thumbnails/17.jpg)
ELS LEFEVER: SENTIMENT ANALYSIS
17
Domain-adaptation: retrain on domain-specific corpus (e.g. “predictable” for a
movie plot /vs/ car’s steering abilities) > supervised/unsupervised
Context-specific: pattern-matching, co-occurrences, machine learning, etc.
E.g. low salary /vs/ low cost
high resolution /vs/ high cost
high quality /vs/ high price
lazy Sunday /vs/ lazy guy
=> problem: need for labeled data, within same domain
![Page 18: PANEL INCREMENTAL LEARNING: HOW SYSTEMS CAN ... · ELS LEFEVER: SENTIMENTANALYSIS 17 Domain-adaptation: retrain on domain-specific corpus (e.g. “predictable” for a movie plot](https://reader035.vdocuments.mx/reader035/viewer/2022070710/5ec742453461017ab76b6522/html5/thumbnails/18.jpg)
ELS LEFEVER: SENTIMENT ANALYSIS
18
How to capture “world knowledge” in an automated way?
The phone (implicit: “good battery”)
=> how to automatically extract the (implicit) aspect + sentiment
Going to the dentist this afternoon. Can’t wait. (implict: “negative
sentiment”)
=> how to model “prototypical sentiment” of phrases?