Instituto de Óptica “Daza de Valdés”

The project “Harnessing Vision Science to Overcome the Critical Limitations of Artificial Neural Networks” has been one of the 5 projects selected in the Fundamentals Program of the BBVA Foundation

Ciencias de la imagen-en

Madrid / February 19, 2024

The call for the Fundamentals Program of the BBVA Foundation has been resolved with the granting of aid of 600,000 euros per project for the development of 5 fundamental and interdisciplinary research in different areas of science.

Among them, the winning project in the category “Mathematics, Statistics, Computer Science, Artificial Intelligence” was the project led by Marcelo Bertalmío, from the CSIC Institute of Optics; and in which Javier Portilla also participates, also a member of the Science group of Image and Vision.

The objective of this Foundations Program is to support exploratory research into central or foundational questions of a scientific field or the intersection of several disciplines.

In this 2022 call, a total of 305 applications have been evaluated and the winning projects will have an execution period of 3 years.

The main objective of the project “Harnessing Vision Science to Overcome the Critical Limitations of Artificial Neural Networks (VIS4NN)” is to develop a new type of artificial neural networks whose behavior is more similar to that of a human observer and that do not have the limitations of current artificial intelligences that make them unable to perform on their own some tasks that are trivial for humans.

Current neural networks are based on the first models of the functioning of vision neurons that, although they are already surpassed by better ones, have allowed artificial neural networks to achieve amazing results in recent years, having multiple applications. in science, industry and society in general.

However, the expectations of the capabilities of these artificial neural networks are colliding with insurmountable limitations that exclude them from being able to perform some tasks such as being able to drive a car completely autonomously.

Of the current limits of artificial neural networks, three are considered the most critical:

  1. Generalization problems. To perform their task well, neural networks need to be trained with a huge data set, which covers all the possibilities of the real world because if they find something unforeseen in their training data set, their chances of getting it right get much worse.
In the ImageNet column, photos of chairs of different types, in the ObjectNet columns photos of chairs in absurd positions
Left column: ImageNet examples. Remaining columns: ObjectNet examples. Neural networks trained on ImageNet suffer a 40-45% performance drop when tested on ObjectNet, showing the generalization problems of artificial neural networks. / Institute of Optics.
  1. Susceptibility to adversarial attacks. It’s easy to make the neural network fail if you know how it works.
4 apparently identical photos that are classified as other objects, for example an African crocodile
The original banana image (top left) is altered by small, intentional perturbations, so that an artificial neural network misclassifies it as a baseball, an electric drill, or an African crocodile, depending on the perturbation. / Institute of Optics
  1. Amount of training data and associated energy costs. The amount of energy that must be consumed to train the increasingly better neural networks increases exponentially and is becoming one more problem to add to global warming.
Graph with a descending line that indicates the error of the AIs and with data on the contamination that is necessary to achieve it
Extrapolation of the current trend suggests that the amount of data, and associated computational resources, required to train an ANN to have less than 5% error in recognizing objects in the ImageNet database would be so large that it would lead to the same CO2 emission that New York City generates in a month. / Institute of Optics
For all these reasons, the financing of this important Fundamentals Program will be used to design new components that serve for a new generation of neural networks that use very recent results and techniques in vision science that go beyond the standard model.
These new neural networks will require much less data to train and will be simpler, reducing CO2 emissions while eliminating the limitations of the previous model.

IO-CSIC Communication

Related news