Download as PDF, PPTX
![7/12/2023 9
Cartoonist Zach Weinersmith, Science: Abridged Beyond the Point of Usefulness, 2017
Deep learning
[…] Now DL works on
everything, except for
small data, shifted data,
noisy data, artificially
twisted data, deep stuffs,
exact stuffs, abstract
stuffs, causal stuffs,
symbolic stuffs, thinking
stuffs, and stuffs that no
one knows how they work
like consciousness.](https://image.slidesharecdn.com/deep-analytics-v2-231207055151-d4f94efd/75/Deep-analytics-via-learning-to-reason-9-2048.jpg)
![Rodney Brooks’
prediction (1/2018)
[In 2023, there is an] emergence of the
generally agreed upon "next big thing" in AI
beyond deep learning.
As of 11/2022, there is no news yet!
http://rodneybrooks.com/my-dated-predictions/](https://image.slidesharecdn.com/deep-analytics-v2-231207055151-d4f94efd/75/Deep-analytics-via-learning-to-reason-10-2048.jpg)
![Why (not) neural reasoning?
Reasoning is not necessarily achieved by
making logical inferences
There is a continuity between [algebraically
rich inference] and [connecting together
trainable learning systems]
Central to reasoning is composition rules to
guide the combinations of modules to address
new tasks
→ Neural networks are a plausible candidate!
7/12/2023 18
“When we observe a visual scene,
when we hear a complex sentence, we
are able to explain in formal terms the
relation of the objects in the scene, or
the precise meaning of the sentence
components. However, there is no
evidence that such a formal analysis
necessarily takes place: we see a
scene, we hear a sentence, and we just
know what they mean. This suggests
the existence of a middle layer,
already a form of reasoning, but not
yet formal or logical.”
Bottou, Léon. "From machine learning to machine
reasoning." Machine learning 94.2 (2014): 133-149.
Leon Bottou](https://image.slidesharecdn.com/deep-analytics-v2-231207055151-d4f94efd/75/Deep-analytics-via-learning-to-reason-18-2048.jpg)
![“[…] one cannot hope to
understand reasoning without
understanding the memory
processes […]”
(Thompson and Feeney, 2014)
7/12/2023 22
Learning to remember](https://image.slidesharecdn.com/deep-analytics-v2-231207055151-d4f94efd/75/Deep-analytics-via-learning-to-reason-22-2048.jpg)
![Noisy, dynamic data: Problem
38
Unified multimodal representation
[Events/features streams] [Sensory streams]
Cross-channel deliberative
reasoning
Anomaly detection
Events detection
Prediction/forecast
Generation
[Downstream tasks]
Dist. events assoc.
Decoder
[Memory]
?](https://image.slidesharecdn.com/deep-analytics-v2-231207055151-d4f94efd/75/Deep-analytics-via-learning-to-reason-38-2048.jpg)
![Noisy, dynamic data: System
39
Unified multimodal representation
[Events/features streams] [Sensory streams]
Cross-channel deliberative
reasoning
Anomaly detection
Events detection
Prediction/forecast
Generation
[Downstream tasks]
Dist. events assoc.
Decoder
[Memory]](https://image.slidesharecdn.com/deep-analytics-v2-231207055151-d4f94efd/75/Deep-analytics-via-learning-to-reason-39-2048.jpg)
Uploaded byDeakin University
Deep analytics via learning to reason
The document discusses advancements in machine learning and AI, highlighting the limitations of current deep learning methods, particularly in reasoning capabilities and generalization. It outlines various approaches to enhancing machine reasoning, including implicit and explicit program synthesis, and emphasizes the importance of integrating memory processes with reasoning. The document also explores applications in healthcare and drug discovery, proposing a Turing machine model for healthcare processes.
More Related Content
Similar to Deep analytics via learning to reason
Deep analytics via learning to reason
- 1. via Learning toreason A/Prof Truyen Tran Deep analytics
- 2.
- 3. Agenda • Where weare • Learning to reason • Applications 7/12/2023 3
- 4.
- 5.
- 6. 7/12/2023 6 convolution -- motifdetection 3 sequencing time gaps/transfer phrase/admission 1 embedding 2 word vector medical record visits/admissions time gap ? prediction point output max-pooling prediction 4 5 record vector Verbalize() – everything as a sentence Nguyen, Phuoc, Truyen Tran, Nilmini Wickramasinghe, and Svetha Venkatesh. Deepr: a convolutional net for medical records." IEEE journal of biomedical and health informatics 21, no. 1 (2016): 22-30.
- 7. 10 years on … •Things have changed • Better supply of data scientists, esp. fuelled by deep learning revolution • Much better understanding of values, processes, and toolsets (MLOps, AutoML) • We still need • Data-driven cultures 7/12/2023 7 https://hbr.org/2022/07/is-data-scientist-still-the-sexiest-job-of-the-21st-century
- 8.
- 9. 7/12/2023 9 Cartoonist ZachWeinersmith, Science: Abridged Beyond the Point of Usefulness, 2017 Deep learning […] Now DL works on everything, except for small data, shifted data, noisy data, artificially twisted data, deep stuffs, exact stuffs, abstract stuffs, causal stuffs, symbolic stuffs, thinking stuffs, and stuffs that no one knows how they work like consciousness.
- 10. Rodney Brooks’ prediction (1/2018) [In2023, there is an] emergence of the generally agreed upon "next big thing" in AI beyond deep learning. As of 11/2022, there is no news yet! http://rodneybrooks.com/my-dated-predictions/
- 11. Agenda • Where weare • Learning to reason • Applications 7/12/2023 11
- 12. What is missingin deep learning? • Modern neural networks are good at interpolating → Data hungry to cover all variations and smooth local manifolds → Little systematic generalization (novel combinations) • Lack of human-perceived reasoning capability • Lack of logical inference • Lack of natural mechanism to incorporate prior knowledge, e.g., common sense • No built-in causal mechanisms 7/12/2023 12
- 13. What is missing: System2 capability • Decoupling from perception/representation (which deep learning does well) • Holds hypothetical thought • Enabling mental travels & imagination. • Slow. Deliberative. Conscious. • Needs working memory. • Size is not essential. Its attentional control is. 7/12/2023 13
- 14. Machine reasoning Reasoning isconcerned with arriving at a deduction about a new combination of circumstances. Reasoning is to deduce new knowledge from previously acquired knowledge in response to a query. 7/12/2023 14 Leslie Valiant Leon Bottou
- 15. (Classical AI) Can weinduce from data a deduction mechanism? (Bayesian AI) Can we infer an inference model from data? (Connectionist AI) Can we design a neural network that learns to reason from (little) demonstration data? 7/12/2023 15 Image credit: MV times
- 16. Hypotheses Reasoning as just- in-timeprogram synthesis. It employs conditional computation. Reasoning is recursive, e.g., mental travel. 7/12/2023 16
- 17. Figure credit: JonathanHui 1980s-2000s: Probabilistic reasoning with Belief Propagation (BP) • BP iteratively computes “beliefs” of unobserved variables based on evidences from observed variables. • Known result in 2001-2003: BP minimises Bethe free-energy minimization. • → Reasoning as belief refinement • But requires full specification of potential functions and network structure! 7/12/2023 17 Heskes, Tom. "Stable fixed points of loopy belief propagation are local minima of the bethe free energy." Advances in neural information processing systems. 2003.
- 18. Why (not) neuralreasoning? Reasoning is not necessarily achieved by making logical inferences There is a continuity between [algebraically rich inference] and [connecting together trainable learning systems] Central to reasoning is composition rules to guide the combinations of modules to address new tasks → Neural networks are a plausible candidate! 7/12/2023 18 “When we observe a visual scene, when we hear a complex sentence, we are able to explain in formal terms the relation of the objects in the scene, or the precise meaning of the sentence components. However, there is no evidence that such a formal analysis necessarily takes place: we see a scene, we hear a sentence, and we just know what they mean. This suggests the existence of a middle layer, already a form of reasoning, but not yet formal or logical.” Bottou, Léon. "From machine learning to machine reasoning." Machine learning 94.2 (2014): 133-149. Leon Bottou
- 19. Two approaches toneural reasoning • Implicit chaining of predicates through recurrence: • Step-wise query-specific attention to relevant concepts & relations. • Iterative concept refinement & combination, e.g., through a working memory. • Answer is computed from the last memory state & question embedding. • Explicit program synthesis: • There is a set of modules, each performs an pre-defined operation. • Question is parse into a symbolic program. • The program is implemented as a computational graph constructed by chaining separate modules. • The program is executed to compute an answer. 7/12/2023 19
- 20. Learning to reason •Learning is to improve itself by experiencing ~ acquiring knowledge & skills • Learning to reason is to improve the ability to decide if a knowledge base entails a predicate. • E.g., given a video f, determines if the person with the hat turns before singing. • Recall the hypotheses: • Reasoning as just-in-time program synthesis. • It employs conditional computation. 7/12/2023 20 Khardon, Roni, and Dan Roth. "Learning to reason." Journal of the ACM (JACM) 44.5 (1997): 697-725. (Dan Roth; ACM Fellow; IJCAI John McCarthy Award)
- 21. Practical setting: (query,database,answer) triplets •Classification: Query = what is this? Database = data. • Regression: Query = how much? Database = data. • QA: Query = NLP question. Database = context/image/text. • Multi-task learning: Query = task ID. Database = data. • Zero-shot learning: Query = task description. Database = data. • Drug-protein binding: Query = drug. Database = protein. • Recommender system: Query = User (or item). Database = inventories (or user base); 7/12/2023 21
- 22. “[…] one cannothope to understand reasoning without understanding the memory processes […]” (Thompson and Feeney, 2014) 7/12/2023 22 Learning to remember
- 23. Memory is essentialto computation and intelligence • Memory is the ability to store, retain and recall information • Brain memory stores items, events and high-level structures • Computer memory stores data and temporary variables • Reasoning capability is correlated with memory capacity 23 Slide credit: Hung Le Heit, Evan, and Brett K. Hayes. "Predicting reasoning from memory." Journal of Experimental Psychology: General 140, no. 1 (2011): 76.
- 24. Memory taxonomy basedon memory content 24 • Objects, events, items, variables, entities Item Memory • Relationships, structures, graphs Relational Memory • Programs, functions, procedures, how-to knowledge Program Memory Slide credit: Hung Le
- 25. Trainable Turing machines Turingmachines are hypothetical devices with infinite memory that can compute all possible computable programs The quest: Learn a Turing machine from data, then use it to solve everything else. Image credit: arstechnica
- 26. Neural Turing machine(NTM) (simulating a differentiable Turing machine) • A controller that takes input/output and talks to an external memory module. • Memory has read/write operations. • The main issue is where to write, and how to update the memory state. • All operations are differentiable. Source: rylanschaeffer.github.io
- 27. Failures of item-onlymemory for reasoning • Relational representation is NOT stored → Can’t reuse later in the chain • A single memory of items and relations → Can’t understand how relational reasoning occurs • The memory-memory relationship is coarse since it is represented as either dot product, or weighted sum. 7/12/2023 27
- 28. Self-attentive associative memories(SAM) Learning relations automatically over time 7/12/2023 28
- 29. Multi-step reasoning overgraphs 7/12/2023 29
- 30.
- 31. NUTM: Learn toselect program (neural weight) via program attention 31 Le, Hung, Truyen Tran, and Svetha Venkatesh. "Neural Stored-program Memory." In International Conference on Learning Representations. 2019. Slide credit: Hung Le Neural stored-program memory (NSM) stores key (the address) and values (the weight) The weight is selected and loaded to the controller of NTM The stored NTM weights and the weight of the NUTM is learnt end-to-end by backpropagation
- 32. System 1: Intuitive System 1: Intuitive System1: Intuitive • Fast • Implicit/automatic • Pattern recognition • Multiple System 2: Analytical • Slow • Deliberate/rational • Careful analysis • Single, sequential Single Image credit: VectorStock | Wikimedia Perception Theory of mind Recursive reasoning Facts Semantics Events and relations Working space Memory A next generation AI architecture @A2I2
- 33. Agenda • Where weare • Learning to reason • Applications 7/12/2023 33
- 34. Electronic medical records Threeinteracting processes: Disease progression Interventions & care processes Recording rules 7/12/2023 34 Source: medicalbillingcodings.org
- 35. Hypothesis: Healthcare is Turing computational • Healthcareprocesses as executable computer program obeying hidden “grammars” • The “grammars” are learnable through observational data 7/12/2023 35 https://www.pinterest.com.au/pin/14073817561987932/ http://workingmodeloftheworld.com/Turing-Machine
- 36. Towards a differentiableTuring machine for health 7/12/2023 36 #REF: Hung Le, Truyen Tran, and Svetha Venkatesh. “Dual Control Memory Augmented Neural Networks for Treatment Recommendations”, PAKDD18. Illness memory Future trajectory Past trajectory
- 37.
- 38. Noisy, dynamic data:Problem 38 Unified multimodal representation [Events/features streams] [Sensory streams] Cross-channel deliberative reasoning Anomaly detection Events detection Prediction/forecast Generation [Downstream tasks] Dist. events assoc. Decoder [Memory] ?
- 39. Noisy, dynamic data:System 39 Unified multimodal representation [Events/features streams] [Sensory streams] Cross-channel deliberative reasoning Anomaly detection Events detection Prediction/forecast Generation [Downstream tasks] Dist. events assoc. Decoder [Memory]
- 40.
- 41. 41 Reasoning Qualitative spatial reasoning Relational, temporal inference Commonsense Objectrecognition Scene graphs Computer Vision Natural Language Processing Machine learning Testbed: Visual QA Parsing Symbol binding Systematic generalisation Learning to classify entailment Unsupervised learning Reinforcement learning Program synthesis Action graphs Event detection Object discovery
- 42. 42 Language-binding Object GraphNetwork for VQA Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran, “Dynamic Language Binding in Relational Visual Reasoning”, IJCAI’20.
- 43. Visual QA inaction
- 44. A general-purpose neuralreasoning unit for spatio-temporal objects (OSTR)
- 45. OSTR in action– Video QA Dang, Long Hoang, et al. "Hierarchical Object-oriented Spatio-Temporal Reasoning for Video Question Answering." IJCAI (2021).
- 46. Video Dialog: Conversation Objects inSpace-Time 12/7/2023 46 • Data = video + dialog + question • Training signal = answer • Model = recurrent reasoning over space-time & dialog turn Hoang-Anh Pham, Thao Le, Vuong Le, Tu Minh Phuong, Truyen Tran. "Unsupervised Attention Priors for Visual Question Answering." ECCV’22
- 47. Drug-target binding asQA • Context: Binding targets (e.g., RNA/protein sequence, or 3D structures), as a set, sequence, or graph. • Query: Drug (e.g., SMILES string, or molecular graph) • Answer: Affinity, binding sites, modulating effects 7/12/2023 47 Nguyen, Tri Minh, Thin Nguyen, Thao Minh Le, and Truyen Tran. "Gefa: early fusion approach in drug-target affinity prediction." IEEE/ACM transactions on computational biology and bioinformatics 19, no. 2 (2021): 718-728.
- 48. Chemical reaction prediction:Reinforcement learning + relational reasoning 48 #REF: Do, K., Tran, T., & Venkatesh, S. (2019, July). Graph transformation policy network for chemical reaction prediction. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (pp. 750-760). ACM. Image credit: britannica The method takes graph morphism as the core. Reinforcement learning is employed to explore the graph morphism dynamics.
- 49. The road ahead Imagesource: pobble365
- 50. Achieving systematic generalization •Refers to the ability to work robustly with new combinations with zero probability in training data. • E.g., if we understand ‘John loves Mary’, then we can also understand ‘Mary loves John’, but machine may fail due to zero probability of the latter if not done properly. • Current DL has a major problem with it. • This is not new: Has been argued for 30+ years! • Much research is needed on multiple fronts (e.g., syntax, indirection, datasets, measuring) 7/12/2023 50 Bahdanau, Dzmitry, et al. "Systematic generalization: what is required and can it be learned?." arXiv preprint arXiv:1811.12889 (2018). Fodor, Jerry A., and Zenon W. Pylyshyn. "Connectionism and cognitive architecture: A critical analysis." Cognition 28.1-2 (1988): 3-71.
- 51. In search forbasic neural operators for reasoning • Basics: • Neuron as feature detector → Sensor, filter • Computational graph → Circuit • Skip-connection → Short circuit • Essentials • Multiplicative gates → AND gate, Transistor, Resistor • Attention mechanism → SWITCH gate • Memory + forgetting → Capacitor + leakage • Compositionality → Modular design • .. 7/12/2023 51 Photo credit: Nicola Asuni
- 52. Searching for reasoningprior: Attention 52 Thao Le, Vuong Le, Sunil Gupta, Svetha Venkatesh, Truyen Tran. "Unsupervised Attention Priors for Visual Question Answering." WACV’23.
- 53. Attention priors 53 Thao Le,Vuong Le, Sunil Gupta, Svetha Venkatesh, Truyen Tran. "Unsupervised Attention Priors for Visual Question Answering." WACV’23
- 54. 54 Analogy operator: Structure-MappingTheory Systematicity Principle: A predicate that belongs to a mappable system of mutually interconnecting relationships is more likely to be imported into the target than is an isolated predicate. Solar system Distance Attractive force Revolves around Color Temperature Hydrogen atom Distance Attractive force Revolves around Color Temperature
- 55. 55 Indirection Layer (InLay) •Concrete data representation is viewed as a complete graph with weighted edges. • The indirection operator maps this graph to a symbolic graph with the same weight edges, however the vertices are fixed and trainable. • This symbolic graph is propagated and the updated node features are indirection representations
- 56. 7/12/2023 56 Thank youTruyen Tran @truyenoz truyentran.github.io truyen.tran@deakin.edu.au letdataspeak.blogspot.com goo.gl/3jJ1O0 linkedin.com/in/truyen-tran We’re hiring PhD & Postdocs truyen.tran@deakin.edu.au https://truyentran.github.io/scholarship.html Tutorial neuralreasoning.github.io
- 57. The reasoning team@ 7/12/2023 57 A/Prof Truyen Tran Dr Thao Le Dr Hung Le Mr Long Dang Mr Hoang-Anh Pham Mr Kha Pham Dr Dung Nguyen