Neuroscientists Recreate Monkey Vision with Tiny AI
Researchers at Cold Spring Harbor Laboratory have developed an AI model small enough to be sent in an email that can reportedly recreate monkey vision. This work in neuroscience contrasts with the massive, supercomputer-trained models typically used in the pursuit of human-like artificial intelligence. The approach aims to better understand how biological brains actually work.
- The model was created by first training a large deep neural network with 60 million parameters to predict the neural responses of macaque monkeys to images, and then shrinking it 5,000-fold. - This dramatic compression was achieved using machine learning techniques known as knowledge distillation and pruning, resulting in a final model with only about 10,000 parameters. - The research is led by Assistant Professor Benjamin Cowley at Cold Spring Harbor Laboratory, whose group focuses on creating compact, interpretable models to understand the brain's computations. - The model's architecture, called a VOneNet, incorporates a "VOneBlock" that uses a biologically-constrained Gabor filter bank to simulate the primate's primary visual cortex (V1). - Unlike many large-scale AI models that are computationally expensive and often brittle, this biologically-inspired approach results in a model that is more robust to image perturbations and adversarial attacks. - This research is part of a broader movement in "Green AI" that emphasizes creating smaller, more energy-efficient models, in contrast to the trend of ever-larger networks that consume vast amounts of power. - By analyzing the compressed model, researchers can more easily understand the step-by-step computations a neuron performs, a task that is nearly impossible with massive, opaque "black box" models. - The ultimate goal of this approach is not just to replicate vision, but to create highly predictive and explainable models that reveal the fundamental principles of biological intelligence.
