Podcast
Questions and Answers
Backpropagation. Is backpropagation considered biologically plausible? If so, why, if not, why not?
Backpropagation. Is backpropagation considered biologically plausible? If so, why, if not, why not?
- No. In biological brains, backpropagation would lead to the vanishing gradient problem, in which gradients would go to zero before reaching the necessary depth in the network.
- Yes. Like many AI principles, it is expressly inspired by known biological feedback connections.
- No. Backpropagation requires identical feedback connections for all feedforward connections. (correct)
- Yes. This is the role played by dopamine in the mammalian brain.
Grid Coding. We can use fMRI to find grid-like coding in the human brain, because:
Grid Coding. We can use fMRI to find grid-like coding in the human brain, because:
- We can’t. As explained in class, grid cells are remapped from environment to environment, so this cannot be done in a reasonable fMRI experiment.
- Depending on your direction (modulo 60 degrees) in a space, you will either traverse the activations or silences of all grid cells. So, we analyze fMRI responses contingent on movement direction. (correct)
- We can use sophisticated machine learning techniques to find the signatures of grid-like coding, especially in entorhinal cortex.
- When running along a hexagonal grid in a virtual reality environment, grid cells will respond periodically (at 4 Hz) whenever you cross the humps of a grid and this fast periodicity is what fMRI picks up.
Divisive Normalization population receptive field modeling. There are multiple parameters in the DN
pRF model that represent constants in the numerator and denominator in its division:
Divisive Normalization population receptive field modeling. There are multiple parameters in the DN pRF model that represent constants in the numerator and denominator in its division:
- Activation constant (B), in the numerator, determines the amount of response compression.
- Normalization constant (D), in the numerator, determines the amount of surround inhibition.
- Normalization constant (D), in the denominator, determines the amount of response compression. Activation constant (B), in the numerator, determines the amount of surround inhibition. (correct)
- Normalization constant (B), in the denominator, determines the amount of surround inhibition.
About divisive normalization: DN allows responses to be scaled to some background, for example it can rescale local neural
responses to light based on the mean light level across the retina
About divisive normalization: DN allows responses to be scaled to some background, for example it can rescale local neural responses to light based on the mean light level across the retina
About divisive normalization: DN can create winner-take-all suppression between neurons that represent different visual
orientations, for example.
About divisive normalization: DN can create winner-take-all suppression between neurons that represent different visual orientations, for example.
About divisize normalization: DN can implement the temporal difference learning algorithm for reinforcement learning, but fails to
capture the dynamics of the Rescorla-Wagner model
About divisize normalization: DN can implement the temporal difference learning algorithm for reinforcement learning, but fails to capture the dynamics of the Rescorla-Wagner model
About divisive normalization: DN is one of the basic ingredients of recurrent neural network architectures, without which
backpropagation-through-time cannot be performed.
About divisive normalization: DN is one of the basic ingredients of recurrent neural network architectures, without which backpropagation-through-time cannot be performed.
Single-layer perceptron. There is a specific operation that a single-layer perceptron cannot perform.
Which?
Single-layer perceptron. There is a specific operation that a single-layer perceptron cannot perform. Which?
Grid Coding. How do responses in Entorhinal cortex (EC) lead to place cell responses in Hippocampus?
Grid Coding. How do responses in Entorhinal cortex (EC) lead to place cell responses in Hippocampus?
Replay. A population of place cells:
Replay. A population of place cells:
Study Notes
Backpropagation and Biological Plausibility
- Backpropagation's biological plausibility is a topic of debate, with some arguing for or against it.
Grid Coding
- fMRI can be used to find grid-like coding in the human brain.
- Grid coding in the Entorhinal cortex (EC) leads to place cell responses in the Hippocampus.
Divisive Normalization
- DN is a mechanism that allows responses to be scaled to a background, such as rescaling local neural responses to light based on the mean light level across the retina.
- DN can create winner-take-all suppression between neurons that represent different visual orientations.
- DN can implement the temporal difference learning algorithm for reinforcement learning, but fails to capture the dynamics of the Rescorla-Wagner model.
- DN is a crucial component of recurrent neural network architectures, enabling backpropagation-through-time.
Single-Layer Perceptron
- There is a specific operation that a single-layer perceptron cannot perform.
Replay
- A population of place cells is involved in the replay process.
Studying That Suits You
Use AI to generate personalized quizzes and flashcards to suit your learning preferences.
Description
Explore the concept of backpropagation and whether it is considered biologically plausible. Learn about the reasons why backpropagation might be seen as biologically plausible or not.
