A deep learning framework for neuroscience
Key Insights
- 3 essential components to specify for artificial systems to do cognitive modelling:
- learning objective/goal (objective (loss) functions, measures the performance of the network)
- learning rules (recipe for updating the synaptic weights)
- network architecture (flow of information)
- One of our key claims is that, even though we must infer them, objective functions are an attainable part of a complete theory of how the architectures or learning rules help to achieve a computational goal.
- starting with the right kind of top-down theoretical framing is important
Inductive Bias
- Inductive biases are a means of embedding such prior knowledge into an optimization system.
- Deep learning works so well, in part, because it uses appropriate inductive biases for the AI Set particularly hierarchical architectures
- by focusing on ANN designs with inductive biases that are useful for the AI Set—so we suspect that it will also be crucial to the development of a deep learning framework for systems neuroscience to focus on how a given animal might solve tasks in its appropriate Brain Set.
- it is worth noting that (i) many species, especially humans, develop slowly with large quantities of experiential data and (ii) deep networks can work well in low-data regimes if they have good inductive biases46. For example, deep networks can learn how to learn quickly
What is the credit assignment problem?
Whenever an algorithm learns there is a ∆F (F = objective function). The difference between current and old performance. So Weights W change.
Gradient-based algorithms aim to choose the option with the most improvement in the desired direction, increasing F.
But which weights need to be updated?
How much reward or punishment should one neuron or synapse get?
The problem with backprob:
- biologically implausible
- too linear
- symmetric feedback
- distinct forward and backward passes of information
- synapses also give direct and immediate feedback to the neuron pre-signal neuron
The problem with random feedback weights:
- algorithms with random feedback weights are highly biased
- the weights that needs to be biased might not be due to randomness
- the feedback does not take the same path as the forward pass
- random assignments might increase increment wrong weights and worsens the algorithm
The paper
see also
Type:
Tags:
Status:
Location:
Created: 30-10-24 17:02
611 📠Machine Learning
Source
Richards, B.A., Lillicrap, T.P., Beaudoin, P. et al. A deep learning framework for neuroscience. Nat Neurosci 22, 1761–1770 (2019). https://doi.org/10.1038/s41593-019-0520-2