Loading...

Modeling and Testing Three-Class Object Recognition and Decision-Making with Comparison of Results

Abbasi Larki, Mahyar | 2025

89 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 58627 (04)
  4. University: Sharif University of Technology
  5. Department: Physics
  6. Advisor(s): Moghimi, Saman; Ebrahimpour, Reza
  7. Abstract:
  8. Object recognition is a fundamental capability in humans and animals that plays a key role in their interaction with the environment. Most existing computational models in this area either neglect the decision stage or employ classical machine-learning models that are not biologically realistic. Moreover, these models are typically limited to binary tasks, whereas decision making in the real world often requires choosing among multiple object categories. In this thesis I introduce a neurocomputational model that simulates the object-recognition process from representation to decision. The model is built by combining a three-class spiking convolutional neural network with a three-population dynamical decision-making model. In the object-recognition component, natural images are encoded as temporal spike-trains and learned in an unsupervised manner using a spike-timing-dependent homeostatic learning rule. In the decision component, three neuronal populations compete with one another, and the population exhibiting the greatest activity determines the final category via an accumulation-to-threshold mechanism. To validate the model, a psychophysical experiment was designed in which participants classified images of three animals — “cow”, “dog”, and “horse” — under different levels of noise; in addition to their choices, participants’ confidence and reaction times were recorded. For further validation of the model’s performance, it was also evaluated on a separate dataset of images of “apple”, “pear”, and “tomato”. Comparison of the model’s results with human data showed that the patterns of change in accuracy, reaction time, and decision confidence were largely similar in both. As noise increased, accuracy decreased in both human and model data, and a strong inverse correlation between decision time and reported confidence was observed. By extending binary decision frameworks to multiclass tasks and providing a biologically plausible model, this research takes a step toward a better understanding of the neural mechanisms of decision making in the brain and may inspire the design of neuromorphic systems — systems whose structure and function are inspired by the human brain and that are optimized for efficient, robust processing of ambiguous inputs.
  9. Keywords:
  10. Spiking Deep Convolutional Neural Networks ; Unsupervised Homeostatic Learning ; Computational Cognitive Neuroscience ; Temporal Object Recognition ; Recurrent Attractor Decision-Making Model ; Three-Class Decision Making ; Decision Confidence

 Digital Object List

 Bookmark

No TOC