Researchers at the Information Sciences Institute at USC Viterbi are developing algorithms that teach machines to learn without human supervision.
“Generally speaking, machine learning is the science of teaching machines to behave like humans,” said Mohammad Rostami, research lead at USC Viterbi’s Information Sciences Institute (ISI). Teaching machines to learn without human supervision is the subject of his latest paper, “Overcoming Conceptual Shifts in Domain Recognition Settings with Integrated Internal Distributions,” presented at the Artificial Intelligence Conference in Washington DC in February. It will be presented at the 37th AAAI Conference on Intelligence. July 14, 2023.
Rostami explains how machine learning generally works: The problem we encounter is that the knowledge the machine acquires is limited to the dataset used for training. Furthermore, the data sets used for training are often unavailable after the training process is complete.
A challenge as a result? When a machine receives inputs that differ significantly from the data it was trained on, it becomes confused and stops behaving like humans.
Bulldog or Shih Tzu or something else?
Rostami gave an example. When training a machine to classify dogs, its knowledge is limited to the samples it was trained on. If there is a new category of dog that is not included in the training sample, the machine cannot learn that it is a new type of dog. “
Interestingly, humans are better than machines. When humans are given something to classify, they adjust and learn what the new category is given just a few samples of a new category (i.e., a new dog breed). Rostami said:
Classification in the Face of Conceptual Change
Often it’s not about learning entirely new categories, but about being able to adjust to changes in existing ones.
If a machine learns a category during training and then undergoes some changes (i.e. adding new subcategories) over time, Rostami suggests that in his work the machine learns or extends that category concept. I hope to be able to. (that is, to include new subcategories).
A change in the nature of a category is known as a “concept shift.” The concept of what a category is changes over time. Rostami gave another example of spam folders.
He explains: It is trained to identify spam using specific features. For example, if an email is not personalized, it is more likely spam. “
Unfortunately, spammers are aware of these models and are constantly adding new features to trick them and prevent emails from being classified as spam.
Rostami continues: This is a time dependent definition. The concept is the same, there is the concept of “spam”, but over time the definition and details of the concept change. That is the concept shift. “
new training methods
In his paper, Rostami developed a method for training machine learning models to address these issues.
Rostami’s method does not rely on the original training data, which is not always available. His co-author and ISI chief scientist, Aram Galstyan, explains: ”
This allows the model to retain what it learned in the initial training phase and adapt and learn new categories and subcategories over time.
It’s also important not to forget the original training data and what you learned from it. This is a big problem in machine learning. Galstyan explains: This is known as catastrophic oblivion,” Galustian said.
According to Galstyan, the approach developed in this paper: Therefore, our model does not forget the old model. ”
what’s next?
Rostami and Galstyan are happy with the results, especially since they don’t rely on the availability of source data. Galstyan said:
Rostami and Galstyan plan to continue work on this concept and apply the proposed method to real-world problems.
But first, Rostami will present his research and findings at the 37th AAAI Conference on Artificial Intelligence. His AAAI conference, run by the largest professional organization in the field, aims to promote artificial intelligence research and scientific exchanges among his AI researchers, practitioners, scientists and engineers in related fields. and This year’s conference acceptance rate was 19.6%.
final highlight
In addition to presenting this paper, Rostami was selected for the new faculty highlight speaker program at AAAI ’23. Rostami, who will become a faculty member at USC in July 2021, will give his 30-minute talk about his research to date and his vision for the future of AI. The program is highly competitive and typically employs less than 15 people, based on past research promise and impact (such as publications in leading forums, citations, awards, or systems deployed) and future plans. of new faculty members. .
Original: Are machines smarter than a 6 year old?
Than: University of Southern California Viterbi School of Engineering
