Attempts to improve physics instruction suggest that there is a fundamental barrier to the human learning of physics. We argue that the new capabilities of artificial intelligence justify a reconsideration not of how we teach physics but to whom we teach physics.

The struggle to teach physics dates back at least to the nineteenth century,1 and remains a losing fight.2 Researchers have tried changing student motivation,3 simplifying course content,4 and developing curricular materials more appealing to students,5 but no significant improvement in student test scores has resulted. This lack of dramatic progress is universally ascribed to the wide variety of student learning types and preparation. The typical difficulties posed by human characteristics have been discussed by Wynken et al.,6 whose results on “pair instruction” are presented in Fig. 1. Proposals to reduce statistical noise through student cloning7 have run into the expected barriers.

Fig. 1.

Results of the “WBN” test of pair instruction. A subset of university physics students at Barber College, Cardinal College, and Faber College were given special recitation sections with pair instruction training. A similar study is ongoing at Wossamotta U. For the pair instruction subset, results were compiled of Δg, the deviations of the average course grade from the average grade for the entire class. These were divided by the standard deviation for the distribution of course grades and plotted as a function of the number Np of pair instruction problems the selected students were given over the course of the semester. The solid curve shows the theoretical expectations of the WBN theory. The dashed curve shows the improved fit of the WBN 25-parameter theory. Of note are the low grades, relative to the class, when students were given more than 60 pair instruction problems. This steep drop-off has been attributed to a growing exasperation of the students with pair instruction, and perhaps to the conflicts that started developing between paired students.

Fig. 1.

Results of the “WBN” test of pair instruction. A subset of university physics students at Barber College, Cardinal College, and Faber College were given special recitation sections with pair instruction training. A similar study is ongoing at Wossamotta U. For the pair instruction subset, results were compiled of Δg, the deviations of the average course grade from the average grade for the entire class. These were divided by the standard deviation for the distribution of course grades and plotted as a function of the number Np of pair instruction problems the selected students were given over the course of the semester. The solid curve shows the theoretical expectations of the WBN theory. The dashed curve shows the improved fit of the WBN 25-parameter theory. Of note are the low grades, relative to the class, when students were given more than 60 pair instruction problems. This steep drop-off has been attributed to a growing exasperation of the students with pair instruction, and perhaps to the conflicts that started developing between paired students.

Close modal

In this paper, we propose an innovative view of instruction in introductory physics, and a solution arising from this view. The focus is on “university physics,” the calculus-based physics introduction designed to bar from engineering careers students who might pose a danger in such careers.

We start with the observation that the failure to improve instruction is, in itself, a very important insight. In an era with a spectacular progress in a wide variety of fields, the efforts of physicists have proved to be wanting in only two areas: quantum gravity and teaching.

The key to this may very well lie in the insights published in 1993 by physicist Cromer.8 He marshals historical data to bolster his argument that physics is not a normal human activity. What sort of activity is university physics then? We first note that it is highly algorithmic. Students learn patterns of problem solving. At Crenshaw-Mellon University,9 in fact, simple computer programs have been developed to recognize and solve the dry-sliding-friction-block-on-tilted-plane, ballistics, and pendulum problems that constitute almost all of university physics.

As the next element of our argument, we note the nature of the typical university physics lecture. The instructor's laptop runs through a set of power point-style slides, while the students at their seats copy this information into their own laptops. The question arises as to the role of the student in this process. There seems, in fact, to be little or no role.

Now, we come to the central question underlying our innovative view: Are we not teaching the wrong entities? “Deep learning” is a very modern version of computer neural networking in which computers are “trained” with a large number of training problems. In tests, the computers then show an apparent intelligence for solving similar new problems. A brief consideration will make clear to any instructor the parallel to teaching university physics. In particular, the students solve a large number of training problems (homework). But, unlike deeply learning computers, the students do not appear to improve with training.10 

With the cooperation of Prof. T. Nodd, we were given data on student performance in the three sections of Faber College's fall 2015 university physics course. In that course, students were given 90 homework problems and took three exams, including the final exam, with a total of 13 problems.

The homework problems were then given as “training” problems to a standard, publicly available, three layered deep-learning program, implemented on a high end desktop workstation. The deep learning program (which was enrolled in the course as D. Plurnur) was then given the problems on the course exams. The machine was given the same time, a total of 5 h, on the three exams, but finished early, needing only 17 ms.

The performance of D. Plurnur was slightly better than that of the human students, though not by a statistically significant margin. The results of the next step, however, were much more encouraging. D. Plurnur was given all the problems in the text as training, along with every problem in two similar textbooks, and the complete set of problems from a publicly accessible compilation of 104 such problems. After this extended training, D. Plurnur did perfectly on all the Faber exams, and achieved an almost perfect score on the Force Concept Inventory.11,12

We wanted to compare this performance with that of Faber students who had solved a large number of training problems, but no reward could be found that would persuade any of the Faber students to attempt more than 90 problems.

Physics is the foundational science. Its disappearance from institutions of higher learning would presage the withering of all academic science and technology. It is important, therefore, that physics instruction continues, but its continuation is threatened. Two thirds of post-secondary institutions are supported by state legislatures, and these legislatures, justifiably, want accountability. As the word suggests, this means counting something, something like credit hours. They are not equipped to judge nor interested in judging the details of instruction, only in the efficiency with which credit hours are generated.

As a boost to efficiency, no other innovation can compare to the shift from the human instructional client to the computing machine client. Very large numbers of instructional clients can be enrolled, and all will get the highest grade possible. Brick and mortar costs will be almost nonexistent, since remote training of computing machines has none of the issues that have been found in remote instruction of human students.

There are countless secondary ways in which this shift will lighten the loads at our overburdened institutions. Protests will be unknown. Trigger warnings will be unnecessary, and, barring remarkable advances in artificial intelligence, sexual harassment will cease to be a problem.

The authors thank S. Candlestickmaker for the pioneering work in this very important branch of physics.

1.
Valerie K.
Otero
and
David E.
Meltzer
, “
100 years of attempts to transform physics education
,”
Phys. Teach.
54
,
523
527
(
2016
).
2.
Lawrence A.
Coleman
,
Donald F.
Holcomb
, and
John S.
Rigden
, “
The introductory university physics project 1987–1995: What has it accomplished?
,”
Am. J. Phys.
66
,
124
137
(
1998
).
3.
Priscilla
Laws
, “
Sweatshop physics
,”
Phys. Tutor
22
,
23
27
(
1997
).
4.
R. S.
Tottle
, “
Newton's laws: Do we really need three?
,”
J. Phys. Simp.
78
,
1
4
(
2012
);
R. S.
Tottle
, “
Kepler's laws: Do we really need three?
,”
J. Phys. Simp.
92
,
22
26
(
2014
);
R. S.
Tottle
, “
Thermodynamic laws: Do we really need three?
,”
J. Phys. Simp.
93
,
6
12
(
2015
);
R. S.
Tottle
, “
Priscilla laws: Do we really need one?
,”
J. Phys. Simp.
94
,
24
40
(
2016
).
5.
Elmak
Dermot
and
Alvin U.
Flinn
, “
A course in the physics of Facebook
,”
Phys. Curric.
1
,
1
13
(
2014
).
6.
Wallace
Wynken
,
I.
Blynken
, and
Tod
Nodd
, “
Expanding the use of pair instruction
,”
Natl. J. Phys.
80
,
376
381
(
2012
).
7.
J.
Fast
and
A.
Modest
, “
Proposal
,”
J. Stu. Clone
6
,
21
53
(
1998
).
8.
Alan
Cromer
,
Uncommon Sense
(
Oxford U.P.
,
New York
,
1993
).
9.
Currently Uber University.
10.
T.
Test
, “
Searching for a signal
,”
J. PER Stat.
128
,
145
623
(
2012
).
11.
D.
Hestenes
,
M.
Wells
, and
G.
Swackhamer
, “
Force concept inventory
,”
Phys. Teach.
30
,
141
158
(
1992
).
12.

D. Plurnur argued that the one answer that was graded wrong was, in fact, correct. The disagreement is being studied.