2. Where is ML at now?
1
What are the key
developments?
2
Directions in Machine Learning (ML)
Working well
Not so good
A good place to
start to find
trends
3. ML Timeline
1
1950s 1980s 2010s
Rules driven/ formal logic/
Rigid top down Reasoning
Data driven approaches
4. ML Timeline
1
1950s 1980s 2010s
Rules driven/ expert systems/
Symbolic Reasoning
Data driven approaches
Human generated knowledge Data Derived knowledge
Programming Training
5. Growth of Machine learning
Improved
Algorithms
Computers are
extremely good at
spotting patterns in
data – beyond human
capability
More
Data
More
Computing
Power
For well bounded tasks
Supervised learning
From labelled
Unsupervised learning
Unlabelled data
8. ML Snags
Machine learning system (trained CNNs) to
identify pneumonia from chest x-rays
dropped in accuracy when used in hospitals
outside of the one it was trained
Difficulty identifying the specific variables
driving predictions
(The machine as able to work out which
hospital a scan had come from by analysing
small metal tokens placed in the corner of
scans, which differed across hospitals)
ML models will find patterns, but not always the ones we think of
or want
Explainability is crucial
9. ML Snags
o 2019 National Institute for Standards and Technology in
US tested 200 facial recognition algorithms, finding that
many were significantly less accurate at identifying
black faces than white ones
o 2017 Amazon abandons its recruitment system which
favours male CVs. Circular problem trained on CVs of
previous successful applications, male etc.
ML models are only as intelligent as the data they are trained on
Bias in the data
10. ML Snags
ML models are only as intelligent as the data they are trained on
Lack of contextual knowledge
“Jake’s birthday party, for his fifth birthday; He’s happy, because kids love their
birthday. He’s about to make a wish and blow out the candles”
Hat
Face
Cake
Emotion = Happy
11. Researchers have already demonstrated how
to fool an AI system into misreading a stop
sign, by carefully positioning stickers on it1.
Self driving model mistakes motorbike for
bobsleigh with parachute
Facial-recognition systems can be deceived by
sticking a printed pattern on glasses or hats
(or painting dots – security issue)
Lack of (top down) reasoning – which humans employ all the time
ML models lack “common sense”
ML Snags
12. Easy for a toddler; Harder for an ML algorithm;
14. March 13th 2020: Netherlands Court legislate that
SYRI an AI powered system designed to detect tax
and social benefit fraud is unlawful.
Works off big data from various government
departments and detects anomalous fraudulent
patterns.
Protection of right to private and family life:
violates Art 8 of the European Convention on
Human Rights
Ethical and privacy considerations
ML Snags
ML models may be relying on private data
15. ML Snags
Training data does not cover all cases (and is expensive)
Challenges developing Amazon Go”Stores - “Walk
in, Pick up, Leave”
Sensors/ cameras capture data about what
happening and ML models determine what the
customer has done
Edge cases:
Overlapping people?
Family groups?
Friends?
Pickup and putting down?
+ ……
“Incompleteness” of data should be known - “what are we not training on”?
Companies planning to exploit deep learning need large amounts of data
16. ML Snags
Machine learning models lose intelligence over time
Concept Drift: Often incremental - e.g. Email spam, fraud detection mechanism
Sometimes abrupt – e.g.
- DNNs for flight terrorism
with new one way flight
patterns in COVID
- Facial recognition with
masks
Knowledge limitations as a result of incomplete data must be known
17. The common trend across the themes is
- extending beyond the training data/
- reducing reliance on expensive hard to get data
18. Trend - hybrid models
Combining two camps: Machine learning & Logical reasoning
Combining data driven machine learning with
top down rules based reasoning for more
powerful, wider learning
Example: The neuro symbolic concept learner
From MIT, IBM and DeepMind (2019)
“We propose the Neuro-Symbolic Concept
Learner (NS-CL), a model that learns visual
concepts, words, and semantic parsing of
sentences without explicit supervision on any of
them; instead, our model learns by simply
looking at images and reading paired questions
and answers.!
19. Trend – pre-existing knowledge
Combining machine learning & pre existing knowledge sources
“The time of having to have labelled data for every class you want to identify is over”
Instead of having to train for all classes – expand machine models with external knowledge bases
to allow models to identify new unseen classes
Zero Shot ; One Shot ; Few shot
Several active research projects in
Generalised zero shot learning:
Example:
Identifying previously unseen unknown images by using zero shot learning with Word2Vec
knowledge space
Using zero shot learning for identifying scenes in video
20. Trend – Greater Reliance on Synthetic Data
Artificially generated data that mimics the real data in terms of essential parameters, univariate and
multivariate distributions, cross-correlations between the variables and so on. Synthetic data can be
created algorithmically
Good for visual /audio data
Amazon Go Stores – Used graphics s/w to create virtual
shoppers and unusual or hard to find scenarios
Self driving cars - trained using high fidelity simulations
Other scenarios:
• Disaster prediction/ recovery models
• Wind farms
• Environmental models
Simulation s/w provides flexible, “free”,
private labels – for hard to acquire data cases
21. Trend – Synthetic data
Concept: The generative network generates candidates while
the discriminative network evaluates them
Generative Adversarial networks
Invented by Ian Goodfellow in 2014
GANs generate synethetic data that resemble real world
data. They consist of unsupervised learning algorithms
that use a supervised loss as part of the training.
Can be used for a variety of data types- and data
conversions.
Active area of research
Our use: to generate fake audio
22. Trend – Mass generation & reuse of knowledge
Knowledge generation
- Knowledge generation factories – 3rd party data preparation market worth 1.5billion dollars in 2019,
- Third party services – e.g. Mechanical turk
Owning and selling this knowledge will be key
- Amazon Go Stores- now licencinsg its technology
- Why reinvent the wheel? E.g. ModelPlay host trained ML models for free or paid use
Research examples for core models – e.g. ResNet and VGG for image processing
- Chinese company MadaCode has 10K freelance
employees;
23. Trend
End of the black box: Models must have transparent
decision making (and “what if” on feature input)
Legal
Usable
e.g. medical domain
Optimisation
By Data scientists
Explainable AI (XAI) is now a front and centre area of research in
machine learning ; Visualisations/ evaluation methods/ widening out
beyond image models. Particularly important for deep learning models
Two new demonstrators in CeADAR on way:
(1) Explainable AI for a deep learning model for structure data (using
Layer wise relevance propogation (LRP), LIME and SHAP
(2) Explainable AI for text
See SOTA documents!
Ethical
For Trust
Commercial
For end customers
X