Zoom In An Introduction to Circuits

By studying the connections between neurons, we can find meaningful algorithms in the weights of neural networks.

Three Speculative Claims about Neural Networks

1. Claim 1: Features Features are the fundamental unit of neural networks. They correspond to directions. 1 These features can be rigorously studied and understood.
2. Claim 2: Circuits Features are connected by weights, forming circuits. 2 These circuits can also be rigorously studied and understood.
3. Claim 3: Universality Analogous features and circuits form across models and tasks.

To support Claim 1 (features) they suggest there are some neurons corresponding to curves (around 60 pixels and at a specific angle) like below. There are some related shapes that use curves but do not count as one.

Understanding Features

curves

How are they sure that these neurons do curve detection?

1. feature visualization, dataset examples, synthetic examples 2.when rotating a sample, curve detections stop firing and curve detections corresponding to the next angle starts firing
2. looking at weights suggests they detect curves

1. next layers features that uses these needs curves (like circle need curves)
2. make these circuits by hand (setting weights) and the behavior is very similar to the neurons.

High-Low Frequency Detectors

these features detect high frequency in one side and low frequency in the other side probably to detect edges and background (out of focus and low frequency) from foreground (high frequency)

Dog feature

Detects dogs from any angle. used feature vis and dataset samples to prove.

Understanding Circuits

curves

earlier curves which are smaller act as tangents of the longer curves. Same sided curves boost the curves in the next layer, and opposite sided curves inhibited next layer curve detectors.

Dog detectors

There are two separate paths in inception for detecting dog heads. The first one detects left oriented heads, while the second one detects right oriented heads.

Then by combining them, the network detects heads in any pose.

car detection

the same story here, car is composed of car wheels, windows, car body.

But after this some interesting things happen. The network decides to move the car to superposition.