Can brain network do BackProp
Created: 01 Mar 2023, 10:06 AM | Modified: =dateformat(this.file.mtime,"dd MMM yyyy, hh:mm a")
Tags: knowledge, KnowledgeSharing
Philosphy - what we believe will eventually affect the type of research that we do
- e.g. Bengio does bio inspired research, Hinton doesn’t have anything related to the human brain
Do weights exist in the brain? In the neuron?
- Yes - potential difference
- It is directed
Biological neural network - the brain has interconnected neurons - but neurons are not all the same type of neuron - there are different types
Assigning the credits - backprop is currently the most efficient way to do it, there are possibly other ways but current research does not have anything fruitful
- This is the “learning”
- How does the brain then assign these credits to each neuron?
- It is not really backprop
- Neurons don’t know the downstream weights
- No evidence for separate forward and backward pass
- It is not really backprop
Forward gradient descent (FGD)
- Getting gradients without backprop
- Calculate slope of vector
- Is this the hinton forward forward algorithm paper?