Why a computer cannot become a full-fledged teacher yet
Today you will not surprise anyone with online training programs for software development, foreign languages and other areas of knowledge. Downloading an application and practicing without "live" teachers is a common scenario. Is it possible to create such an ecosystem for human learning that will work completely automatically?
In this article, we will analyze the principles of algorithmic learning and its future. After all, who knows, perhaps in 20 years we will be taught by a full-fledged artificial intelligence, which will know the subject ten times better than any human tutor? Let's look into this further.
What is algorithmic learning?
Algorithmic or programmed learning is instruction using programmed algorithms and without the direct involvement of human teachers.
When a student is diligently working through a computer or a mobile phone, studying English vocabulary in an app, this is a very good example of utilizing algorithmic learning. There is a student, there is a program, and nothing else is needed for the acquisition of knowledge.
The main benefits of learning with apps and programs are:
- Maximum autonomy - You can adjust the difficulty, length and training schedule. You can study anytime, anywhere
- Retention - Simple bits of information presented within the learning program are easy to remember
- Development of logical thinking - The training goes in order: from simple to complex
But these are not all the opportunities that are available to the student. Let’s consider the three further stages of the development of algorithmic learning and take algorithmic learning (e.g. the current educational apps that exist today) as a base.
Stage 1. Fully programmed teaching
Fully programmed teaching was originally described and implemented back in the 70s, so this approach has already accumulated a lot of practical experience.
Representing a completely scripted type of training, the work proceeds only according to a pre-established system of classes. There are simply no other options.
The learning process itself is extremely simple and consists of 3 repetitive stages:
- Processing of educational information. The program gives the student material, explains it in theory and with examples. Consider a case where a student learns English vocabulary in the app and adds a new collection of words to their curriculum. First, the program introduces them to new words and explains their use in different situations - if desired, the student can learn more simply by clicking on an incomprehensible word
- Practical work with material and knowledge testing. During this subsequent stage, the student goes through exercises on new material, which further consolidate knowledge. Next, the program tests the student. Most often, this is a regular multiple-choice test
- Assessment of results and transition to the next stage. To pass each test successfully, one needs to score a predetermined percentage of correct answers. If the test is passed, the system gives the student the next block of information. If not, it recommends that the student goes through the previous block once again and then attempts another test to ascertain mastery
Programmed learning operates without too much data and logical operations. The most complex part of it is the “if, then” action, which has been feasible to accomplish via computers for many decades now. The process itself is completely linear and, essentially, looks like this:
- A? B? C? D
- If D? 80%, then E (go to a new chain)
- If D <80%, then A
This is the simplest type of programmed learning. In some cases, it is really effective, but it is generally impossible to organize a complex learning ecosystem into this extremely structured approach. You can increase logical chains to incredible sizes, but the result will still be insufficient for full mastery of most aspects of human knowledge.
Stage 2. Modular learning and self-learning algorithms
Modular learning gives one much more freedom in process than linear learning. It is actively used today when creating various kinds of systems for training. The entire field of EdTech is, in one way or another, based precisely on modular learning.
All tutorials in the system are divided into modules that are logically related to each other, but formally are independent units. Each module boosts a certain area of knowledge. The order of work on the modules depends on the actions of the student himself - the system, using the assessment of effectiveness, will itself determine what needs to be worked on at a particular moment to maximize results.
It is obvious that modular training is also pre-programmed and all actions taking place within the limits of scripts pre-installed in the system. However, the variability of actions within the system is simply enormous.
Let's look at the example of learning English. This ecosystem consists of several hundred training modules. Among them, for example, are "Study of articles", "Punctuation of complex sentences", "Irregular verbs", and "Common idioms". The number of such modules depends entirely on the scale of the project. The more discrete modules there are, the more individualistic the training can be made.
When working with tutorials, the system analyzes the results by many parameters at once. For each module, the functional parameters are different, but they may overlap. In fact, a huge matrix is created in which the system analyzes hundreds of parameters. During the analysis, the system determines the module or modules that need to be paid attention to right now.
The peculiarity of algorithmic learning is that it does not affect motivation. Of course, it’s possible to add elements of gamification and rewards, which will encourage the student to return to study, but so far it is impossible to track the decline in motivation in this way.
Stage 3. Neural networks and artificial intelligence
A neural network is a special kind of analytical software that works on the same principles as the human brain. In fact, this is a complex algorithm that is able to operate with methods of pattern recognition, discriminant analysis and information clustering.
The main difference between neural networks and standard programmed learning is that neural networks are not programmed, but trained. The system is able to independently identify patterns in disparate information, generalize and classify data. A neural network can analyze information, while programmed systems can only create the illusion of analysis.
Today neural networks are actively tested on systems based on mathematical analysis. For example, in December 2017, DeepMind presented the AlphaZero neural network to the public, which in 24 hours reached a simply colossal level of play, defeating all existing computer chess programs.
The Elo rating is used to evaluate chess players. The maximum rating of Magnus Carlsen, the current world chess champion, is 2863 points. This is a human record. The average rating of a chess master is 2000-2200 points, with grandmasters only slightly higher. At the same time, according to approximate estimates, AlphaZero's Elo rating ranges from 3500 to 5000 points. It's like comparing a world champion to a rookie. Only, now the rookie is a human grandmaster.
Indeed, neural networks perform well in the analysis of mathematical systems and data. However, no one has yet tried to "tie" them to teaching people. The main difficulty is the lack of direct logic in teaching methods. Teaching methods highly dependent on human psychology. There are many subjective features that cannot be expressed in the form of an understandable mathematical system.
Take the following example: The student doesn't like Jennifer Lawrence as an actress. She simply annoys the individual. Meanwhile, the neural network chooses an excerpt from the movie "The Hunger Games", where she plays the main role, as a teaching aid.
The student does not show very good results, because the material is not interesting to them. The neural network is not able to guess why this happened, since this subjective preference does not fit into the principles of mathematical logic. At the same time, a live teacher would be able to quickly figure out the situation, perhaps noting subtle clues such as the body language of the student, and select more relevant material, maintaining motivation and inspiring the learner.
Another reason why neural networks will not teach people soon is the speed of information retrieval. Neural networks are trained independently based on the received data. In mathematical systems, this is simple - the very same AlphaZero simulated millions of games with itself to find the optimal ways of playing. But in learning, the result depends primarily on the people. One lesson is normally at least half an hour. To collect data for analysis requires hundreds of thousands of lessons. It takes years of work for a neural network to reach at least the level that programmed training provides.
Perhaps in the long run artificial intelligence can completely replace teachers and teach concepts faster and better, but this won’t happen soon. The era of modular learning will continue, and it will be so until neural networks reach a qualitatively new level. Nowadays, effective teaching requires a combination of computerized applications and real teachers. This is the optimal way to reach the desired result.