Finding Faults in Autistic and Software Active Inductive Learning

1. Finding Faults in Autistic and Software Active Inductive Learning Boris Galitskyand Igor Shpitsberg Knowledge-Trail Inc. and Rehabilitation Center “Our Sunny World” .

2. Observations • Both real life autistic and machine cognitive systems lack features allowing them to build and use an adequate model of external world. • Both autistic and machine learning systems display deficiencies compared to learning of controls

3. Both machine learning and autistic system have problems understanding intentions of others “Do you want to play here?” “Do you like this?” “Are you hungry?” “Am I hungry?” “Does she want to play with you?” “Do I want to know this?” Relevant: understand and react adequately Irrelevant: do not understand => no reaction

4. Questions • How can finding commonalities in autistic and machine learning deficiencies help better understand them? • How can observing autistic cognition help designing better active machine learning systems? • How can observing certain problems in designing machine learning systems help better understanding problems of autistic learning? Both most machines and most children with autism cannot deceive. To teach them, we need an explicit definition

5. Objective • Describe an autistic learning mechanism from the computational learning standpoint • Design a machine learning system which would display autistic behavior in its auto-development • Observe which features of autistic learning can be reproduced • Help correction of autistic cognition and learning

6. Learning is deterministic, inductive, active and reward-based • We use deterministic learning model to avoid uncertainty features and maintain as simple model as possible • We use inductivelearning to obey a clear cause-effect structure, following the traditional inductive schema. The commonality in stimuli is assumed to cause an effect (a target feature) • Learning isactive, since the system itself selects the new elements of training set • Learning is reward-based, so each correct stimulus recognition problem solved is rewarded. Incoming stimulus are selected from the real world, and the learning system itself chooses the best stimuli to recognize

7. Active learning loop

8. Sensory perception of children with autism • It is well known that sensory perception of children with autism is rather peculiar • A vast number of children with autism successfully ignore one kind of sensory stimulus and totally intolerable to the others • They can form simple and complex sequences from various subjects and action, but at the same time refuse to reproduce simple schemata suggested by their teachers

9. Hypersensitivity • We hypothesize that a root cause of autistic cognition is hypersensitivity to input stimuli • Many studies of the dis-ontogenesis and the peculiarities of the development of children with autism confirm it at the earlier stages of ontogenesis(Baron-Cohen ‎2004; Marco 2011) • It becomes clear that the development of an adequate sensory system at consecutive development steps by an autistic child is impossible

10. From hyper-sensitivity to failed cognition • Autistic learning system is initially adequate but hyper-sensitive • It deviates stronger and stronger from both development of control children and adequate machine learning systems • Instead of collecting richer and richer stimuli of the real world, it learns to ignore them and substitute with auto stimulation • Attempting to recognize real stimuli, such learning system receives negative reward We simulate such behavior computationally and explain how initial hyper-sensitivity leads to a number of limitations of learning system, inherent to autistic learning.

11. Hyper-sensitive machine learning system If an anomaly detection system is hypersensitive, it becomes dysfunctional If a customer service agent becomes hypersensitive to details, it becomes dysfunctional as well

12. Learning weak instead of strong stimuli • In the efforts to protect themselves from stimuli which are too strong, children with autism develop a mechanism to filter them out • These strong stimuli are mostly more informative than the weak ones • Autistic child picks up weaker stimuli , less informative, but with a higher similarity with each other.

13. Avoiding perception of the real world

14. Image of the mom vsTV commercials • As an example of such stimuli in visual space, let us consider recognition of (1) child’s mother and (2) repetitive TV commercials • Since the perceived image of mother’s face varies more significantly than the perceived image of repetitive TV commercials, the latter is preferred • Image of the mother can be filtered out as being too strong due to its higher variability • It required higher recognition efforts

15. Repetitive stimuli • A partial case of stimuli with high similarity • All children select to use as highly repetitive stimuli as possible as the training set • However autistic children only select most repetitive stimuli and do not proceed beyond them • As a result of this initial problem, children with autism stop exploring human behavior and do not communicate properly with their mothers and other humans

16. Simulate autistic development as a choice of perception mode • a child selectsto recognize humans such as parents and relatives • a child follows an “easier” way of perception, considering only very similar patterns coming as a sequence, such as TV commercials. which requires multimodal perception, classification of rather distinct images in a single pattern, and further emotional and mental development. This child is deprived of mental and emotional development due to his incapability to perceive humans and their mental attitudes

17. Learning from data with high similarity • If a machine learning system is fed with very similar elements of the training set, it will have a problem of recognizing even fairly similar objects to the training ones. • It will be unable to recognize the ones with significant deviation from the elements of the training set • To be rewarded, such learning system would need to find input stimuli which are alike to be able to recognize them. The whole learning capability will be lacking

18. Visual and tactile multi-modal perception

19. Movement and perception of space in autistic development

20. Visual and tactile auto-stimulation

21. Lets proceed to computational learning

22. Generalized active inductive learning procedure with positive and negative cases.

23. Active learning loop

24. Faulty active learning scenarios in the real world • Hypersensitivity of the learning system can be viewed as a high number of features which are mutually correlated, and therefore redundant • The learning algorithm itself can reasonably tackle such situation of overfitting …but the active learning would be selecting training objects which would not adequately cover the real world • To keep being awarded for recognition, the system will at some point stop collecting training objects from external world and start using the existing ones and therefore its proper recognition will not occur. This is essentially an auto-stimulation

25. Our previous studies The Theory of Mindaccount is extended to reflect the computational experience of “teaching” computers to reason about mental attitudes (Galitsky 2000, Galitsky 2005). Various forms of autistic reasoning about action, time, space and probabilities are explored, as well as the prevalence of deductive over inductive, abductive, and analogical forms of reasoning (Galitsky & Goldberg 2003). Training of reasoning about beliefs, desires and intentions is shown to assists the emotional development (Galitsky 2001) A series of interactive rehabilitation software tools is suggested which stimulate various forms of commonsense reasoning, conversation and decision-making in autistic patients (Peterson et al 2004, Galitsky 2002).

26. Conclusions We designed a plausible machine learning system which shows two forms of behavior: • normal mode, where new features from the real world form the training dataset and form the basis for its proper recognition • autistic faulty mode, where the active learning evolves to the set of irrelevant features and although the learning sessions occur, the system is not capable of recognizing the real world

27. Conclusions: features of autistic cognition Given the operational learning system: once it becomes hyper-sensitive in an active learning mode, it displays the number of features inherent to autistic cognition: • Avoiding strong and informative stimuli • Broken multi-modal links • Auto-stimulation • Mixing important and unimportant features • Ability to learn only from a training set with very high similarity / uniformity / repetition