Anti-Learning Adam Kowalczyk Statistical Machine Learning NICTA, Canberra 1 National ICT Australia Limited is funded and supported by:
Overview Anti-learning –Elevated XOR Natural data –Predicting Chemo-Radio-Therapy (CRT) response for Oesophageal Cancer –Classifying Aryl Hydrocarbon Receptor genes Synthetic data –High dimensional mimicry Conclusions Appendix: A Theory of Anti-learning –Perfect anti-learning –Class-symmetric kernels
Definition of anti-learning Training accuracy Random guessing accuracy Off-training accuracy Off-training accuracy Systematically:
Anti-learning in Low Dimensions y x z +1 -1
Anti-Learning Learning
Evaluation Measure Area under Receiver Operating Characteristic (AROC) f fθ False Positive True Positive AROC( f )
Learning and anti-learning mode of supervised classification TP FN AROC FN AROC FN TP + + Learning Anti- learning AR OC Test Training Random: AROC = 0.5 ?
Anti-learning in Cancer Genomics
From Oesophageal Cancer to machine learning challenge
Learning and anti-learning mode of supervised classification TP FN AROC FN AROC FN TP + + Learning Anti-learning AROC Test Training Random: AROC = 0.5
Anti-learning in Classification of Genes in Yeast
KDD’02 task: identification of Aryl Hydrocarbon Receptor genes (AHR data)
Anti-learning in AHR-data set from KDD Cup 2002 Average of 100 trials; random splits: training: test = 66% : 34%
KDD Cup 2002 Yeast Gene Regulation Prediction Task Vogel- AI Insight - change - change or control Single class SVM 38/84 training examples 1.3/2.8% of data used in ~14,000 dimensions
Anti-learning in High Dimensional Approximation (Mimicry)
Paradox of High Dimensional Mimicry high dimensional features If detection is based of large number of features, the imposters are samples from a distribution with the marginals perfectly matching distribution of individual features for a finite genuine sample, then imposters are be perfectly detectable by ML-filters in the anti-learning mode
Mimicry in High Dimensional Spaces
Quality of mimicry Average of independent test for of 50 repeats d = 1000 d = 5000 = | n E | / | n X |
Formal result :
Proof idea 1: Geometry of the mimicry data Key Lemma:
Proof idea 1: Geometry of the mimicry data
Proof idea 2:
Proof idea 3:kernel matrix
Proof idea 4
Theory of anti-learning
Hadamard Matrix
CS-kernels
Perfect learning/anti-learning for CS-kernels Kowalczyk & Chapelle, ALT’ 05 False positive True positive Test ROC S-T Train ROC T 1 1
Perfect learning/anti-learning for CS-kernels Kowalczyk & Chapelle, ALT’ 05
Perfect learning/anti-learning for CS-kernels
Perfect anti-learning theorem Kowalczyk & Smola, Conditions for Anti-Learning
Anti-learning in classification of Hadamard dataset Kowalczyk & Smola, Conditions for Anti-Learning
AHR data set from KDD Cup’02 Kowalczyk, Smola, submitted Kowalczyk & Smola, Conditions for Anti-Learning
From Anti-learning to learning Class Symmetric CS– kernel case Kowalczyk & Chapelle, ALT’ 05
Perfect anti-learning : i.i.d. a learning curve n = 100, n Rand = 1000 random AROC: mean ± std n samples i.i.d. samples from the perfect anti-learning-set S More is not necessarily better!
Conclusions Statistics and machine learning are indispensable components of forthcoming revolution in medical diagnostics based on genomic profiling High dimensionality of the data poses new challenges pushing statistical techniques into uncharted waters Challenges of biological data can stimulate novel directions of machine learning research
Acknowledgements Telstra –Bhavani Raskutti Peter MacCallum Cancer Centre –David Bowtell –Coung Duong –Wayne Phillips MPI –Cheng Soon Ong –Olivier Chapelle NICTA –Alex Smola