1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
|
You can start by clicking the "start" button.
You will not know what to fill in
if you don't have the correct way of thinking.
Sometimes I don’t have any ideas.
To get a high score,
testers need to reason abstractly
from the information they extracted
from small samples.
A high score cannot be obtained by just doing a lot of exercises.
Towards general artificial intelligence.
Exploring cognitive intelligence with human-level intelligence quotient.
"Artificial intelligence" has always been a window for humans
to explore the boundaries of their capabilities.
In recent years, significant progress in artificial intelligence
represented by deep learning has been made at the perception level,
but there is still a long way for existing models
to achieve intelligence with general human-level cognitive capabilities.
Research has shown that in the case of determining
whether two figures are alike,
primates like capuchin monkeys can do it successfully.
This indicates that animals have an innate cognitive architecture
that allows them to find generic paradigms
for solving problems from small data.
These advantages of cognitive framing are particularly evident in humans.
For slightly more complex geometric problems, for instance,
the Amazonian indigene group in the rainforest
can still solve them easily.
However, the deep learning foundation model represented by Transformer
is dwarfed in similar tests
not only does the model require a large amount of labeled data for training,
but its ultimate performance cannot be comparable to that of humans.
Intelligence levels are generally measured
based on intelligence quotients,
or "IQ" as it is often called.
Psychologists have created a series of tests
to numerically quantify IQ
and have found that IQ level has a high correlation
with human achievement.
Among these tests,
a representative one is Raven's Progressive Matrices.
The following question is an example.
This example is complicated at first glance,
which has only 8 pictures and the shapes of objects are different.
However, a closer analysis shows that
the objects in each row are all dark gray, light gray and black,
and the size of the objects in each picture is basically the same.
Thus, it is not difficult to find the correct answer.
In the case of Odd-One-Out, on the other hand,
subjects are required to pick an outlier data point from several examples.
For example, in the next question,
only the third picture has a dark black hexagon.
For traditional perceptual intelligences,
we need to provide thousands of examples
for the machine to learn the concept of a cat or a dog.
For a cognitive intelligent agent,
however, the machine can abstract the corresponding events from a huge space
from just a few pictures and understand their spatial-temporal-causal relationships.
Exploring models with human cognitive intelligence is
a fundamental research project of the Beijing Institute of General Artificial Intelligence (BIGAI),
in which scholars from BIGAI and UCLA cooperate to address this challenging problem:
how to use small data to understand
spatial-temporal-causal relationships in IQ tests.
After several years of study,
we proposed the Tong-Hui model.
In this summer, we invited students from
top universities in China to
have a competition with our Tong-Hui model.
In the preliminary tests,
we had a rough estimate of the capabilities of the model.
But when faced with truly highly intelligent human opponents,
we were not sure how our model would perform.
All right, I will click the "start" button
to begin the competition.
Students often have a variety of wondrous ideas,
but our program may not have similar thoughts.
Thus, we are not sure
about the result of the competition.
It was quite easy at the beginning,
then it was a little bit tough,
and later I had no idea.
I made 6 to 7 mistakes.
One would make mistakes if he cannot find the correct way of thinking and the hidden pattern,
and he would not know what to fill in.
We needed to spend
a lot of time thinking,
but the machine could quickly
try various solutions in a short period of time.
OK, thank you all.
The Tong Hui model outperformed all the students
and the foundation model represented by Transformer.
The first item in the upper left corner is a pentagon,
the others do not have pentagons.
We have beaten the best students in the country in this task.
Our next step is to provide more robust criteria
for the grading of AI
and to evaluate our general AI systems in a more comprehensive setting.
We were always thinking that
if we ever really created intelligence that
could outperform the world's smartest brains,
we must have discovered some kind of universal algorithm
or even a whole new cognitive architecture.
Perhaps we are already on
the doorstep of general AI right now,
and the success of this competition
will be a further step to it.
|
