The document discusses the evolution of machine reasoning and learning, comparing various cognitive architectures and their limitations from the 1960s to the present. It explores the interplay between learning and reasoning, emphasizing the need for contextual and common-sense reasoning in AI systems, and highlights advancements in visual question answering frameworks. The insights from dual-process theories and neural Turing machines are presented as essential components for developing more effective reasoning capabilities in artificial intelligence.

1. 20/08/2020 1
@FIT, Monash Uni, August 2020
A/Prof Truyen Tran
Vuong Le, Thao Le, Hung Le,
Dung Nguyen, Tin Pham
@truyenoz
truyentran.github.io
truyen.tran@deakin.edu.au
letdataspeak.blogspot.com
goo.gl/3jJ1O0
Machine reasoning
linkedin.com/in/truyen-tran

2. 20/08/2020 2
Dual-process theories
The problem
Neural memories
Agenda
Visual reasoning

3. 20/08/2020 3
Among the most challenging scientific questions of our
time are the corresponding analytic and synthetic
problems:
• How does the brain function?
• Can we design a machine which will simulate a brain?
-- Automata Studies, 1956.

4. 20/08/2020 4
#REF: A Cognitive Architecture Approach to Interactive Task Learning, John E. Laird, University of Michigan

5. 1960s-1990s
 Hand-crafting rules,
domain-specific, logic-
based
 High in reasoning
 Can’t scale.
 Fail on unseen cases.
20/08/2020 5
2020s-2030s
 Learning + reasoning, general
purpose, human-like
 Has contextual and common-
sense reasoning
 Requires less data
 Adapt to change
 Explainable
1990s-present
 Machine learning, general
purpose, statistics-based
 Low in reasoning
 Needs lots of data
 Less adaptive
 Little explanation
Photo credit: DARPA

6. Learning vs Reasoning
Learning is to improve itself by experiencing ~ acquiring
knowledge & skills
Reasoning is to deduce knowledge from previously acquired
knowledge in response to a query (or a cues)
Early theories of intelligence (a) focuses solely on reasoning, (b)
learning can be added separately and later! (Khardon & Roth,
1997).
Learning precedes reasoning or the two interacting?
Can reasoning be learnt? Are learning and reasoning indeed
different facets of the same mental/computational process?
20/08/2020 6
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)

7. Rich Sutton’s Bitter Lesson
20/08/2020 7
“The biggest lesson that can be read from
70 years of AI research is that general
methods that leverage computation are
ultimately the most effective, and by a
large margin. ”
“The two methods that seem to scale
arbitrarily in this way
are search and learning.”
http://www.incompleteideas.net/IncIdeas/BitterLesson.html

8. Bottou’s vision of 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
20/08/2020 8
“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.

9. Learning to reason, formal def
20/08/2020 9
Khardon, Roni, and Dan Roth. "Learning to reason." Journal of the ACM
(JACM) 44.5 (1997): 697-725.
E.g., given a video f, determines if the person with the
hat turns before singing.

10. 10
What color is the thing with the same
size as the blue cylinder?
blue
• The network guessed the
most common color in the
image.
• Linguistic bias.
• Requires multi-step
reasoning: find blue cylinder
➔ locate other object of the
same size ➔ determine its
color (green).
A testbed: Visual QA

11. 11Adapted from (Aditya et al., 2019)
Events,
Activities
Objects, Regions,
Trajectories
Scene Elements: Volumes, 3D
Surfaces
Image Elements: Regions, Edges,
Textures
Knowledge about
shapes
Qualitative spatial
reasoning
Relations between
objects
Commonsense knowledge
Relations between objects
Why Visual QA?

12. 20/08/2020 12
Dual-process theories
The problem
Neural memories
Agenda
Visual reasoning

13. Dual-process system with memories
13
System 1:
Intuitive
System 1:
Intuitive
System 1:
Intuitive
• Fast
• Implicit/automatic
• Pattern recognition
• Multiple
System 2:
Analytical
• Slow
• Deliberate/rational
• Careful analysis
• Single, sequential
Facts
Semantics
Events and relational associations
Working space – temporal buffer
Single

14. System 2
Holds hypothetical thought
Decoupling from representation
Working memory size is not essential. Its
attentional control is.
20/08/2020 14

15. Hypotheses
Reasoning as just-in-time program synthesis.
It employs conditional computation.
Reasoning is recursive, e.g., mental travel.
20/08/2020 15

16. 20/08/2020 16
Dual-process theories
The problem
Neural memories
Agenda
Visual reasoning

17. Visual recognition is a (solvable) special case
17
Image courtesy: https://dcist.com/
What is this?
a cat R-CNN,
YOLO,…
Object recognition Object detection
Where are objects, and what are they?

18. But Image QA is challenging
18
(CLERV, Johnson et al., 2017)
(Q) How many objects are either
small cylinders or metal things?
(Q) Are there an equal number of
large things and metal spheres?
(GQA, Hudson et al., 2019)
(Q) What is the brown animal sitting
inside of?
(Q) Is there a bag to the right of the
green door?

19. And Video QA is even more challenging
19(Data: TGIF-QA)

20. 20
Reasoning
Computer Vision
Natural Language
Processing
Machine
learning
Visual Reasoning
Visual QA is a great
testbed

21. Reasoning over visual modality
• Context:
 Videos/Images: compositional data
 Hierarchy of visual information: objects, actions, attributes,
relations, commonsense knowledge.
 Relation of abstract objects: spatial, temporal, semantic aspects.
• Research questions:
 How to represent object/action relations in images and videos?
 How to reason with object/action relations?
21

22. A simple VQA framework that works surprisingly well
22
Visual
representation
Question
What is the brown
animal sitting inside of?
Textual
representation
Fusion
Predict
answer
Answer
box
CNN
Word embedding + LSTM

23. 20/08/2020 23
A Prelim Dual-System Model
Le, Thao Minh, Vuong Le, Svetha Venkatesh, and Truyen Tran. "Neural
Reasoning, Fast and Slow, for Video Question Answering." IJCNN (2020).

24. Separate reasoning process from perception
24
 Video QA: inherent dynamic nature of visual content over time.
 Recent success in visual reasoning with multi-step inference and
handling of compositionality.
System 1: visual
representation
System 2: High-
level reasoning

25. 20/08/2020 25
Separate reasoning process from perception (2)

26. System 1: Clip-based Relation Network
26
For 𝑘𝑘 = 2,3, … , 𝐾𝐾 where ℎ𝜙𝜙 and
𝑔𝑔𝜃𝜃 are linear transformations
with parameters 𝜙𝜙 and 𝜃𝜃,
respectively, for feature fusion.
Why temporal relations?
• Situate an event/action in relation to
events/actions in the past and
formulate hypotheses on future
events.
• Long-range sequential modeling.

27. System 2: MAC Net
20/08/2020 27
Hudson, Drew A., and Christopher D. Manning. "Compositional attention
networks for machine reasoning." ICLR 2018.

28. Results on SVQA dataset.
28
Model Exist Count
Integer
Comparison
Attribute Comparison Query
All
More Equal Less Color Size Type Dir Shape Color Size Type Dir Shape
SA(S) 51.7 36.3 72.7 54.8 58.6 52.2 53.6 52.7 53.0 52.3 29.0 54.0 55.7 38.1 46.3 43.1
TA-
GRU(T)
54.6 36.6 73.0 57.3 57.7 53.8 53.4 54.8 55.1 52.4 22.0 54.8 55.5 41.7 42.9 44.2
SA+TA 52.0 38.2 74.3 57.7 61.6 56.0 55.9 53.4 57.5 53.0 23.4 63.3 62.9 43.2 41.7 44.9
CRN+M
AC
72.8 56.7 84.5 71.7 75.9 70.5 76.2 90.7 75.9 57.2 76.1 92.8 91.0 87.4 85.4 75.8

29. Results on TGIF-QA dataset.
29
Model Action Trans. FrameQA Count
ST-TP 62.9 69.4 49.5 4.32
Co-Mem 68.2 74.3 51.5 4.10
PSAC 70.4 76.9 55.7 4.27
CRN+MAC 71.3 78.7 59.2 4.23

30. 20/08/2020 30
A System 1 for Video
Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran, “Hierarchical
conditional relation networks for video question answering”, CVPR'20.
System 1: visual
representation
System 2: High-level
reasoning

31. Conditional Relation Network Unit
31
Problems:
• Lack of a generic mechanism for
modeling the interaction of
multimodal inputs.
• Flaws in temporal relations of
breaking local motion in videos.
Inputs:
• An array of 𝑛𝑛 objects
• Conditioning feature
Output:
• An array of 𝑚𝑚 objects

32. Hierarchical
Conditional Relation
Networks
32

33. Results
33
Model Action Trans. FrameQA Count
ST-TP 62.9 69.4 49.5 4.32
Co-Mem 68.2 74.3 51.5 4.10
PSAC 70.4 76.9 55.7 4.27
HME 73.9 77.8 53.8 4.02
HCRN 75.0 81.4 55.9 3.82
TGIF-QA dataset MSVD-QA and MSRVTT-QA datasets.

34. 20/08/2020 34
A System 2 for Relational Reasoning
Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran, “Dynamic
Language Binding in Relational Visual Reasoning”, IJCAI’20.
System 1: visual
representation
System 2: High-level
reasoning

35. Reasoning with structured representation of
spatial relation
35
• Key insight: Reasoning is chaining of relational predicates
to arrive at a final conclusion
→ Needs to uncover spatial relations, conditioned on query
→ Chaining is query-driven
→ Objects/language needs binding
→ Object semantics is query-dependent
→ Very thing is end-to-end differentiable

36. 36
Language-binding Object Graph Network for VQA

37. 37
Language-binding Object Graph Unit (LOG)

38. 20/08/2020 38

39. Results in CLEVR dataset
39

40. 40
Comparison with the on CLEVR dataset of
different data fractions.
Performance comparison on
VQA v2 subset of long questions.
Model Action
XNM(Objects) 43.4
MAC(ResNet) 40.7
MAC(Objects) 45.5
LOGNet(Objects) 46.8

41. 20/08/2020 41

42. 20/08/2020 42
Dual-process theories
The problem
Neural memoriesVisual reasoning
Agenda

43. “[…] one cannot hope to understand
reasoning without understanding the memory
processes […]”
(Thompson and Feeney, 2014)
20/08/2020 43

44. Dual-process system with memories
44
System 1:
Intuitive
System 1:
Intuitive
System 1:
Intuitive
• Fast
• Implicit/automatic
• Pattern recognition
• Multiple
System 2:
Analytical
• Slow
• Deliberate/rational
• Careful analysis
• Single, sequential
Facts
Semantics
Events and relational associations
Working space – temporal buffer
Single

45. From psychology & cognitive science
Reasoning and memory are studied separately, but …
 Analytic reasoning requires working memory
 Spatial reasoning requires place memory
 Inferential reasoning requires episodic memory
Dual-processes in memory:
 Fast retrieval/recognition (using similarity as criteria)
 Slow item recollection
Dual-processes in reasoning
 Induction as fast (using similarity as criteria)
 Deduction as slow

46. Reasoning as memory
Expert reasoning was enabled by a large long-term memory,
acquired through experience
Working memory for analytic reasoning
 WM capacity predicts IQ
 Higher order cognition = creating & manipulating relations  representation of
premises, temporarily stored in WM
 WM is a system to support information binding to a coordinate system
 Reasoning as deliberative hypothesis testing  memory-retrieval based hypothesis
generation
Knowledge structures support reasoning
Fuzzy-trace theory: gist ( intuition) and verbatim memory (
analytic processing)
Memory is critical for episodic future thinking (mental simulation)
20/08/2020 46

47. 20/08/2020 47
Learning a Turing machine
 Can we learn a (neural)
program that learns to
program from data?
Visual reasoning is a
specific program of two
inputs (visual, linguistic)

48. Neural Turing machine (NTM)
A memory-augmented neural network (MANN)
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

49. MANN for reasoning
Three steps:
 Store data into memory
 Read query, process sequentially, consult memory
 Output answer
Behind the scene:
 Memory contains data & results of intermediate steps
 LOGNet does the same, memory consists of object representations
Drawbacks of current MANNs:
 No memory of controllers  Less modularity and compositionality when
query is complex
 No memory of relations  Much harder to chain predicates.
20/08/2020 49
Source: rylanschaeffer.github.io

50. 20/08/2020 50
Universal Neural Turing Machines
Hung Le, Truyen Tran, Svetha Venkatesh, “Neural stored-program
memory”, ICLR'20.
Hung Le, Truyen Tran, Svetha Venkatesh, “Neurocoder: Learning General-
Purpose Computation Using Stored Neural Programs”, Work in progress.

51. Computing devices vs neural counterparts
FSM (1943) ↔ RNNs (1982)
PDA (1954) ↔ Stack RNN (1993)
TM (1936) ↔ NTM (2014)
The missing piece: A memory to store programs
Neural stored-program memory
UTM/VNA (1936/1945) ↔ NUTM--ours (2019)

52. NUTM = NTM + NSM

53. Question answering (bAbI dataset)
Credit: hexahedria

54. Neurocoder
Flexible programs are essential for intelligence
Cognitive science: Meta-level process controls object-level process
Instance-specific program vs Task-specific program
In QA, program is query-specific
 On-demand program synthesis
Idea: A program to generate programs
Previous works: Fast weights, HyperNetworks, Neural Module
Networks, Adaptive Mixture of Experts
Current work: Programs are generated by combining basis sub-
programs
 The hope: Basis programs span the space of all possible programs
20/08/2020 54

55. Neurocoder (cont.)
20/08/2020 55

56. 20/08/2020 56

57. 20/08/2020 57
Neural relational memory
Hung Le, Truyen Tran, Svetha Venkatesh, “Self-attentive associative
memory”, ICML'20.

58. Failures of current MANNs for reasoning
Relational representationis 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.
20/08/2020 58

59. Self-attentive associative memories (SAM)
Learning relations automatically over time
20/08/2020 59

60. Multi-step reasoning over graphs
20/08/2020 60

61. Reasoning on text
20/08/2020 61

62. 20/08/2020 62
Thank you Truyen Tran
@truyenoz
truyentran.github.io
truyen.tran@deakin.edu.au
letdataspeak.blogspot.com
goo.gl/3jJ1O0
linkedin.com/in/truyen-tran